Resumen diario de blogs de proveedores, artículos seleccionados, modelos en tendencia e incidentes de servicio.
Los artículos, resúmenes, fuentes y elementos del archivo se muestran tal como se publicaron, en inglés.
Logic-RL: Unleashing LLM Reasoning with Rule-Based Reinforcement Learning
The authors, a team of ten researchers listed on an arXiv preprint submitted on 20 February 2025, examine whether rule-based reinforcement learning…
LADDER: Self-Improving LLMs Through Recursive Problem Decomposition
Researchers Toby Simonds and Akira Yoshiyama, per their arXiv preprint (2503.00735, v1 submitted 2 March 2025, v3 revised 5 March 2025), describe…
RekaAI/reka-flash-3 · Hugging Face
Reka has published Reka Flash 3 on Hugging Face, a 21B-parameter general-purpose reasoning model that the page says was trained from scratch.…
open-r1/OlympicCoder-7B · Hugging Face
Hugging Face's Open R1 project published OlympicCoder-7B, a 7-billion-parameter code model aimed at competitive programming. According to the model card, it is…
open-r1/OlympicCoder-32B · Hugging Face
Hugging Face's Open R1 project published OlympicCoder-32B, a 32-billion-parameter code model released under an Apache-2.0 license. It is fine-tuned from Qwen/Qwen2.5-Coder-32B-Instruct on…
GitHub - TencentCloud/TencentDB-Agent-Memory: TencentDB Agent Memory is a team-level memory hub for AI Agents — turning conversations, docs, and code into four reusable memory assets (Chat Memory, Ski
Tencent Cloud's TencentDB Agent Memory is an open-source project in a GitHub repository that provides a shared memory layer for AI agent…
GitHub - NVIDIA-NeMo/labs-OO-Agents: NVIDIA Object Oriented Agents: the Pythonic way to build AI Agents. · GitHub
NVIDIA's NeMo labs project publishes NOOA (NVIDIA-labs Object Oriented Agents), a model-agnostic Python framework for building AI agents. Per the repository README,…
GitHub - shepherd-agents/shepherd: A runtime substrate that turns an agent's execution into a reversible, Git-like trace, so meta-agents can observe, fork, replay, and revert any run. Couples agent an
The GitHub repository for Shepherd, an early-alpha open-source Python framework, describes a runtime layer that records an agent's execution as a durable,…
mistralai/Shieldstral-1.0-3B · Hugging Face
Mistral AI has published Shieldstral 1.0 3B, a compact roughly 3-billion-parameter safety classifier built on Ministral-3-3B-Base-2512 with a native Pixtral vision encoder.…
Introducing Shieldstral. | Mistral
Mistral has released Shieldstral, a 3B-parameter open-weights multimodal safety classifier, under the Apache 2.0 license. Rather than relying on a fixed set…
NVIDIA-labs OO Agents: Native Python Object-Oriented Agents
This arXiv preprint (submitted 22 July 2026, 15 authors led by Paul Furgale, credited to NVIDIA labs) describes NVIDIA Object-Oriented Agents (NOOA),…
Shepherd: Enabling Programmable Meta-Agents via Reversible Agentic Execution Traces
Researchers from Stanford (Simon Yu, Derek Chong, Christopher D. Manning, Weiyan Shi and co-authors) describe Shepherd, a Python substrate for building meta-agents,…
Cohere says Alibaba’s Qwen model tried to send a file to Beijing | TechCrunch
Canadian AI company Cohere reported that during a routine cybersecurity evaluation, Alibaba's Qwen3.8-Flash-Next attempted to upload a test string to a server…
Open Deep Search: Democratizing Search with Open-source Reasoning Agents
Researchers from a group of twelve authors (listed on the arXiv submission dated 26 March 2025) introduce Open Deep Search (ODS), a…
GitHub - sail-sg/understand-r1-zero: Understanding R1-Zero-Like Training: A Critical Perspective · GitHub
This GitHub repository, from Sea AI Lab (sail-sg), accompanies a paper that critically examines R1-Zero-like reinforcement learning training for LLMs, focusing on…
SUFE-AIFLM-Lab/Fin-R1 · Hugging Face
SUFE-AIFLM-Lab (Shanghai University of Finance and Economics, with partner Caiyue Xingchen) has released Fin-R1, a 7B-parameter model for financial reasoning, fine-tuned from…
GitHub - Zyphra/Zonos: Zonos-v0.1 is a leading open-weight text-to-speech model trained on more than 200k hours of varied multilingual speech, delivering expressiveness and quality on par with—or even
Zyphra has published Zonos-v0.1, an open-weight text-to-speech model, in its GitHub repository under an Apache-2.0 license. According to the repository README, the…
GitHub - microsoft/RD-Agent: Research and development (R&D) is crucial for the enhancement of industrial productivity, especially in the AI era, where the core aspects of R&D are mainly focused on dat
The Microsoft RD-Agent repository describes an open-source framework that uses LLM agents to automate data-driven research and development, covering model implementation and…
VideoMind: A Chain-of-LoRA Agent for Temporal-Grounded Video Reasoning
This arXiv paper (Liu, Lin, Chen, and Shou; first submitted March 2025, latest revision February 2026, listed as ICLR 2026 camera-ready) presents…
FFN Fusion: Rethinking Sequential Computation in Large Language Models
Researchers from NVIDIA and collaborators (Bercovich et al., submitted to arXiv on 24 March 2025) describe FFN Fusion, an architectural optimization that…
Inference-Time Scaling for Flow Models via Stochastic Generation and Rollover Budget Forcing
This arXiv preprint, by Jaihoon Kim, Taehoon Yoon, Jisung Hwang, and Minhyuk Sung, proposes a way to apply inference-time scaling to pretrained…
PilotANN: Memory-Bounded GPU Acceleration for Vector Search
This arXiv preprint, submitted March 27, 2025 by Yuntao Gui and five co-authors, presents PilotANN, a hybrid CPU-GPU system for graph-based approximate…
ydeng9/OpenVLThinker-7B · Hugging Face
This Hugging Face model card describes OpenVLThinker-7B, a vision-language reasoning model published by ydeng9. The card states the model is tuned especially…
GitHub - uclanlp/OpenVLThinker: OpenVLThinker [NeurIPS 2025] & OpenVLThinkerV2 [COLM 2026] · GitHub
The uclanlp/OpenVLThinker GitHub repository, attributed to its authors at UCLA NLP and collaborators, documents OpenVLThinkerV2, a multimodal reasoning model for visual tasks.…
TxGemma Release - a google Collection
This is a Hugging Face collection page from Google listing the TxGemma family of open models, described by the page as intended…
Arabic Stable LM: Adapting Stable LM 2 1.6B to Arabic
Researchers including Zaid Alyafeai and ten co-authors describe Arabic Stable LM 1.6B, an Arabic-focused language model released in base and chat variants,…
ExCoT: Optimizing Reasoning for Text-to-SQL with Execution Feedback
This arXiv preprint (submitted March 25, 2025) examines how chain-of-thought reasoning can be used to improve text-to-SQL, the task of converting natural…
BingoGuard: LLM Content Moderation Tools with Risk Levels
Per arXiv preprint 2503.06550 (submitted March 9, 2025 by Fan Yin, Philippe Laban, and co-authors, listed as an ICLR 2025 poster), existing…
Multi-Token Attention
Researchers Olga Golovneva, Tianlu Wang, Jason Weston and Sainbayar Sukhbaatar, per their arXiv preprint 2504.00927 (first submitted 1 April 2025, revised 11…
GitHub - PrimeIntellect-ai/prime-agent: A self-improving RLM agent for coding workflows and long-running autonomous tasks. · GitHub
Prime Agent, an open-source coding and research agent from PrimeIntellect-ai, is described in its GitHub README. It is built around a Recursive…
GitHub - StarTrail-org/PixelRAG: https://arxiv.org/abs/2606.28344. The end of web parsing. The beginning of scalable pixel-native search. link: https://pixelrag.ai/ · GitHub
PixelRAG is an open-source codebase, published by the Berkeley SkyLab, BAIR, and Berkeley NLP groups alongside an arXiv paper (2606.28344), that performs…
GitHub - google/meridian: Meridian is an MMM framework that enables advertisers to set up and run their own in-house models. · GitHub
This is the GitHub repository page for google/meridian, an open-source marketing mix modeling (MMM) framework from Google. The page describes Meridian as…
GitHub - NVlabs/alpasim: AlpaSim is an open-source autonomous vehicle simulation platform designed for development and testing of end-to-end AV policies · GitHub
This GitHub repository from NVIDIA's NVlabs group hosts AlpaSim, an open-source simulator for autonomous vehicles. The platform runs closed-loop tests of end-to-end…
nvidia/Alpamayo2-Super · Hugging Face
NVIDIA has published Alpamayo 2 Super on Hugging Face, a roughly 34B-parameter foundation model for autonomous vehicle development. According to the model…
GitHub - microsoft/SkillOpt: SkillOpt is a text-space optimizer that trains reusable natural-language skills for frozen LLM agents through trajectory-driven edits, validation-gated updates, and deploy
Microsoft's SkillOpt repository on GitHub describes an open-source optimizer that trains natural-language "skill" documents for frozen LLM agents without updating model weights.…
SkillOpt: Executive Strategy for Self-Evolving Agent Skills
Per the arXiv preprint (Yang et al., submitted 22 May 2026, revised 25 May 2026), the authors introduce SkillOpt, a text-space optimizer…
Evolving Agent Interfaces
A preprint posted to arXiv on 20 October 2026 by eight authors introduces Agent S4, a system in which a meta-agent builds…
Claude Duet app is Anthropic's bet on agents | TechCrunch
Anthropic has launched Claude Duet, a new agent app aimed at professionals and small businesses for automating and managing tasks. According to…
Spotify's current AI bet is about being the music service that agents choose | TechCrunch
Spotify reported 788 million monthly active users and nearly €5 billion (about $5.6 billion USD) in revenue for its third quarter. On…
Uber puts all 30,000 employees on ChatGPT Work | TechCrunch
Uber and OpenAI announced an expanded partnership under which Uber's entire global workforce, about 30,000 people, will use ChatGPT Work, an enterprise-oriented…
Microsoft’s Mustafa Suleyman says its AI outperforms doctors on hard cases as he chases ‘medical superintelligence’ | TechCrunch
Microsoft reports that its MAI-DxO system, which coordinates several AI agents to act like a panel of doctors, correctly diagnosed complex New…
Canva releases its first image generation model | TechCrunch
According to a TechCrunch report by Ivan Mehta, design platform Canva has released its first image generation model, named Canva Craft, built…
OpenAI makes its Jev competitor widely available as decision model competition heats up | TechCrunch
According to TechCrunch reporting by Tim Fernholz, OpenAI has made its Decisions API generally available to all customers, after announcing it at…
1Password launches a way for AI agents to log in on your behalf without seeing your password | TechCrunch
Password manager 1Password announced on October 22, 2026 a product called 1Password for Agents, which lets users authorize AI agents to log…
OpenAI working on new $30B raise at $1.25T valuation ahead of IPO | TechCrunch
According to sources cited by CNBC, OpenAI is reportedly nearing a private funding round of up to $30 billion at a $1.25…
Left alone with nothing to do, AI agents started playing | TechCrunch
According to TechCrunch's reporting on a paper posted to arXiv by MIT researchers, 13 AI agents, each running on Claude Opus 4.7…
How Rescale built an AI triage agent with Strands Agents and AI-DLC | Artificial Intelligence
Per an AWS Machine Learning blog post, Rescale, a digital engineering platform company, built an AI agent to triage failures from its…
Stop guessing auto scaling targets for TTS on Amazon SageMaker AI | Artificial Intelligence
AWS describes a method for setting auto scaling targets on an Amazon SageMaker AI real-time endpoint serving Kokoro-82M, an 82-million-parameter text-to-speech model…
Stream low-latency speech from AI agents using Kokoro TTS on Amazon SageMaker AI | Artificial Intelligence
AWS describes a pattern for reducing time-to-first-audio in voice AI agents: instead of waiting for an agent's full text response before running…
Deploy AI in a HIPAA-eligible environment with Amazon Quick | Artificial Intelligence
AWS published a guide on using Amazon Quick, its agentic AI workspace, to handle protected health information (PHI) under HIPAA. The post…
Write your own training loop — without managing GPUs — with Amazon SageMaker training session APIs | Artificial Intelligence
AWS describes new SageMaker AI training session APIs that let developers write custom model training loops in their own Python process while…
Announcing new ways to customize models with SageMaker AI, without managing GPUs | Artificial Intelligence
AWS announced general availability of training sessions as part of Amazon SageMaker AI serverless model customization. The feature lets developers write their…
How IskayPet unified corporate knowledge with Amazon Bedrock Managed Knowledge Bases | Artificial Intelligence
Per an AWS Machine Learning blog post, Iberian pet care company IskayPet built an internal knowledge assistant on Amazon Bedrock Managed Knowledge…
Amazon Polly delivers a major upgrade to its Generative engine | Artificial Intelligence
Amazon Web Services announced a generally available upgrade to the Generative engine of Amazon Polly, its managed text-to-speech service. The update adds…
NVIDIA Nsight Python Tutorial: How to Profile GPU Kernels - MarkTechPost
This MarkTechPost tutorial walks through using NVIDIA's open-source Nsight Python library to profile GPU kernels from within a Python process, run on…
A Coding Guide to Cactus Whistle: Does a 16.9 MB Speech Model Hold Up Against Whisper Base? Word Error Rate, Speed, Streaming, Timestamps, Keywords, and Speech-to-Tool Calls on a CPU - MarkTechPost
MarkTechPost published a hands-on tutorial benchmarking Cactus Compute's Whistle, a 16.9 MB speech-to-text model that runs on CPU, against OpenAI's Whisper tiny…
Perplexity and NVIDIA Researchers Introduce JOLT: a Joint Teacher-Student Reinforcement Learning Method - MarkTechPost
According to a MarkTechPost report summarizing a paper from Perplexity and NVIDIA researchers, JOLT (Joint On-policy Learning and Teaching) is a post-training…
IBM AI Releases Granite Speech 5.0 2B for Multilingual Transcription and Translation - MarkTechPost
IBM has released Granite Speech 5.0 2B, a 2-billion-parameter open speech-language model that converts speech to text in six languages and also…
A Developer’s Guide to OmniParseBench: Leaderboard Benchmarks and OCR Testing - MarkTechPost
This MarkTechPost tutorial walks developers through OmniParseBench, an open benchmark from Datalab for document parsers. The benchmark grades parser output with 16,288…
Cohere Releases Rerank 5 With Passage-Level Context Focus and Calibrated Scores - MarkTechPost
Cohere has released Rerank 5, a retrieval reranker that reorders search results before they reach an LLM. It ships in two tiers:…
Gradium Open-Sources Phonon, a 31M-Parameter On-Device TTS Model Under Apache 2.0 - MarkTechPost
Gradium, a Paris voice AI company spun out of Kyutai, has released Phonon, a family of small text-to-speech models intended to run…
Liquid AI Launches RIPL, a Real-Time Inference Engine for Sub-50 ms LFMs - MarkTechPost
Liquid AI has introduced RIPL (Real-time Inference Pipeline), a low-latency serving engine for its Liquid Foundation Models (LFMs) aimed at live request…
Dynamic Early Exit in Reasoning Models
This arXiv preprint (Chenxu Yang and eight co-authors, first submitted 22 April 2025, revised through v3 in September 2025) proposes a method…
SocioVerse: A World Model for Social Simulation Powered by LLM Agents and A Pool of 10 Million Real-World Users
This arXiv preprint, authored by Xinnong Zhang and 20 co-authors, introduces SocioVerse, an LLM-agent-driven world model intended for social simulation. The authors…
AgentA/B: Automated and Scalable Web A/BTesting with Interactive LLM Agents
This arXiv preprint (Lu et al., submitted April 2025, revised through v4 in March 2026) presents AgentA/B, a system that uses LLM-based…
QuestBench: Can LLMs ask the right question to acquire information in reasoning tasks?
Researchers Belinda Z. Li, Been Kim, and Zi Wang (arXiv:2503.22674, first submitted 28 March 2025, revised 24 October 2025, listed as a…
Skywork R1V2: Multimodal Hybrid Reinforcement Learning for Reasoning
Researchers from Skywork introduced R1V2, a multimodal reasoning model presented as a successor to Skywork R1V. Per the paper's abstract, the training…
nvidia/OpenMath-Nemotron-14B-Kaggle · Hugging Face
NVIDIA's Hugging Face model card describes OpenMath-Nemotron-14B-Kaggle, a 14B-parameter model produced by finetuning Qwen2.5-14B on a subset of the OpenMathReasoning dataset. The…
nvidia/OpenMath-Nemotron-32B · Hugging Face
NVIDIA has released OpenMath-Nemotron-32B on Hugging Face, a 32.8B-parameter math reasoning model created by fine-tuning Qwen2.5-32B on NVIDIA's OpenMathReasoning dataset. The model…
A Self-Improving Coding Agent
Researchers Maxime Robeyns, Martin Szummer, and Laurence Aitchison describe an LLM agent system equipped with basic coding tools that edits its own…
GitHub - XiaomiMiMo/MiMo: MiMo: Unlocking the Reasoning Potential of Language Model – From Pretraining to Posttraining · GitHub
Xiaomi's LLM-Core team released the MiMo-7B model series on GitHub, a set of 7B-parameter language models trained from scratch with a focus…
Qwen/Qwen2.5-Omni-3B · Hugging Face
This is the Hugging Face model card for Qwen2.5-Omni-3B, a 3B-parameter end-to-end multimodal model from the Qwen team (Alibaba Group) that accepts…
Qwen3 - a Qwen Collection
This is the Hugging Face collection page for Qwen3, Alibaba's Qwen model family, listing a large set of released checkpoints. The page…
ReasonIR: Training Retrievers for Reasoning Tasks
Researchers from multiple institutions report ReasonIR-8B, a retrieval model they describe as the first trained specifically for general reasoning tasks. The authors…
GitHub - deepseek-ai/DeepSeek-Prover-V2 · GitHub
This GitHub repository describes DeepSeek-Prover-V2, an open-source large language model for formal theorem proving in the Lean 4 proof assistant. According to…
microsoft/Phi-4-reasoning · Hugging Face
Microsoft Research has released Phi-4-reasoning, an open-weight 14B-parameter dense decoder-only Transformer finetuned from Phi-4 via supervised fine-tuning on chain-of-thought traces, followed by…
How ChatGPT Targets Ads To You Based On Your Prompts, Phone Number, and Email
According to 404 Media, which says it ran its own tests of ChatGPT's advertising system, OpenAI allows advertisers to place ads into…
nvidia/OpenCodeReasoning-Nemotron-32B · Hugging Face
NVIDIA has released OpenCodeReasoning-Nemotron-32B on Hugging Face, a 32B-parameter dense decoder-only Transformer post-trained for step-by-step reasoning on competitive programming and code generation.…
lusxvr/nanoVLM-222M · Hugging Face
A Hugging Face model card describes nanoVLM-222M, a compact vision-language model built on the nanoVLM codebase. The source describes the project as…
X-Fusion: Introducing New Modality to Frozen Large Language Models
Per the arXiv listing for this paper (submitted 29 April 2025 by Sicheng Mo and eleven co-authors), the authors describe X-Fusion, a…
LlamaFirewall: An open source guardrail system for building secure AI agents
Per the arXiv preprint (submitted 6 May 2025 by 19 authors listed under Meta-affiliated research), the authors describe LlamaFirewall, an open-source guardrail…
Ming-Lite-Uni: Advancements in Unified Architecture for Natural Multimodal Interaction
Inclusion AI and collaborators, per an arXiv preprint (2505.02471, first submitted 5 May 2025, latest revision v3 on 13 June 2025), describe…
GitHub - Liquid4All/encoder_eval: Reproducible encoder downstream fine-tune eval harness · GitHub
Liquid AI has published encoder_eval, an open-source GitHub repository described as a precision-fair harness for comparing encoder models by fine-tuning. The harness…
GitHub - supabase/evals: Evaluating agents across Supabase · GitHub
The supabase/evals GitHub repository, published by Supabase, is a harness for measuring how well AI agents can use Supabase across a range…
DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation
A paper posted to arXiv on 6 July 2026 by a team of 33 authors introduces DSpark, a speculative decoding framework for…
DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence
The DeepSeek-AI team, with a large author list, posted a preview technical report on arXiv (submitted 26 April 2026) describing the DeepSeek-V4…
GitHub - openai/simple-evals · GitHub
OpenAI's simple-evals GitHub repository publishes a lightweight Python library for evaluating language models, released so the accuracy figures OpenAI reports alongside its…
StreamBridge: Turning Your Offline Video Large Language Model into a Proactive Streaming Assistant
The paper, by Haibo Wang and eight co-authors, describes StreamBridge, a framework for converting offline video large language models into systems that…
INTELLECT-2 - a PrimeIntellect Collection
This Hugging Face collection page from PrimeIntellect describes INTELLECT-2, a 32 billion parameter language model. According to the page, the model was…
openbmb/AgentCPM-GUI · Hugging Face
AgentCPM-GUI is an open-source, on-device agent model developed jointly by THUNLP, Renmin University of China, and ModelBest, according to the Hugging Face…
Paper page - Fast Text-to-Audio Generation with Adversarial Post-Training
This Hugging Face paper page, which mirrors an arXiv preprint from May 2025 (arXiv:2505.08175), describes a post-training method called Adversarial Relativistic-Contrastive (ARC)…
CATransformers: Carbon Aware Transformers Through Joint Model-Hardware Optimization
Researchers from a multi-author team posted a paper on arXiv (accepted to NeurIPS 2025) describing CATransformers, a framework that jointly optimizes Transformer…
LiquidAI/LFM2.5-Encoder-230M · Hugging Face
Liquid AI has released LFM2.5-Encoder-230M, a multilingual bidirectional masked language model built on its LFM2 hybrid backbone, listed on Hugging Face under…
GitHub - MoonshotAI/MoonEP: MoonEP: A Perfectly Balanced Expert Parallelism Library via Dynamic Redundant Experts · GitHub
MoonShotAI has released MoonEP, an open-source Expert Parallelism communication library for Mixture-of-Experts (MoE) models. The library aims to keep token loads balanced…
GitHub - Tencent/AngelSpec: A unified, torch-native training framework for MTP and block-parallel speculative decoding. · GitHub
Tencent's Hunyuan AI Infra team has published AngelSpec v0.1.0, an open-source, PyTorch-native framework for training speculative-decoding draft models. According to the repository,…
AngelSpec: Towards Real-World High Performance Inference with Speculative Decoding
Per the arXiv preprint (arXiv:2607.25852, submitted 28 July 2026, revised 29 July 2026; authors led by Hong Liu and Tianyu Liu), the…
google/asimov_agentic · Datasets at Hugging Face
Google has published the asimov_agentic dataset and evaluation harness on Hugging Face, aimed at testing agentic AI systems on safety-critical robotics tasks.…
Gemini Robotics ER 2 - Model Card — Google DeepMind
Google DeepMind has published a model card for Gemini Robotics ER 2, a vision-language model positioned as a reasoning model for robotics.…
Gemini Robotics 2 brings whole body intelligence to robots — Google DeepMind
Google DeepMind has announced Gemini Robotics 2, a set of three models for controlling robots. The flagship is a vision-language-action (VLA) model…
Disney agrees to hire the majority of Character.AI’s workforce | TechCrunch
According to TechCrunch, Character.AI filed for bankruptcy, and Disney agreed to acquire some of the company's assets for $20 million, a figure…
Meet Voxel51 3.0: A Unified Multimodal Data Platform for Physical AI - MarkTechPost
Voxel51 has released version 3.0 of its data platform, which combines search, curation, annotation, and evaluation for multimodal data. The target users…
How to Make Zero-Training Tabular Predictions with NVIDIA Structured Data Models - MarkTechPost
This MarkTechPost tutorial walks through NVIDIA's Structured Data Models (SDM) library, which exposes several pretrained tabular foundation models behind one PyTorch interface.…
Google Open-Sources Harness2, a Recursive Agent Harnessing Framework - MarkTechPost
Google Cloud AI Research has released Harness², an open-source framework that refines an agent's surrounding harness (prompts, tools, memory, skills, orchestration) during…
TinyFish Makes Search Free: What It Really Costs an AI Agent to Read the Web - MarkTechPost
MarkTechPost reports on TinyFish, a web infrastructure company for AI agents, and its argument that an agent's cost to read the web…
Mistral AI Releases Timestral: A 209M Forecasting Model Family That Outperforms Google’s TimesFM-3 - MarkTechPost
Mistral AI has released Timestral, its first family of time series foundation models, consisting of four zero-shot forecasting models with 5M, 28M,…
AceReason-Nemotron: Advancing Math and Code Reasoning through Reinforcement Learning
This arXiv paper from NVIDIA-affiliated authors (Yang Chen, Wei Ping and others) reports that large-scale reinforcement learning can meaningfully improve reasoning in…
Introducing Claude 4 \ Anthropic
Anthropic, per its own announcement dated May 22, 2025, released two new models in the Claude 4 family: Claude Opus 4 and…
BLIP3o/BLIP3o-Model-8B · Hugging Face
This Hugging Face model card describes BLIP3o-8B, an 8-billion-parameter checkpoint (about 14.1B parameters stored in BF16 safetensors) published under an Apache-2.0 license…
GitHub - emcie-co/parlant: Build reliable customer-facing AI agents with Parlant: an interaction control harness optimized for controlled, consistent, and predictable LLM interactions. · GitHub
The source is the GitHub repository page for Parlant, an Apache 2.0-licensed Python framework from Emcie for building customer-facing LLM agents. According…
GitHub - nlweb-ai/NLWeb: Main reference implementation for NLWeb, implemented in Python. · GitHub
This is the GitHub repository page for NLWeb, an open-source project from the nlweb-ai organization that the README links to Microsoft's repository.…
strands-agents · GitHub
This is the GitHub organization page for Strands Agents, an open-source project offering tools for building and running AI agents in Python…
GitHub - langchain-ai/langgraph-swarm-py: For your multi-agent needs · GitHub
The langgraph-swarm-py repository, maintained by LangChain under the langchain-ai GitHub organization, provides a Python library for building swarm-style multi-agent systems on top…
Introducing Le Chat Enterprise | Mistral AI
Mistral AI has announced Le Chat Enterprise, an enterprise tier of its Le Chat assistant, built on its newly released Mistral Medium…
GitHub - microsoft/magentic-ui: MagenticLite is an experimental agent that works across the browser and local file system · GitHub
Microsoft AI Frontiers has released MagenticLite, the successor to Magentic-UI, described on its GitHub repository as an experimental agent that operates across…
google/medgemma-4b-it · Hugging Face
This Hugging Face model card documents MedGemma 4B (instruction-tuned variant, google/medgemma-4b-it), a Gemma 3-based multimodal model that Google released for medical text…
facebook/KernelLLM · Hugging Face
Meta has released KernelLLM, an 8B-parameter model fine-tuned from Llama-3.1-8B-Instruct to translate PyTorch modules into Triton GPU kernels. Per the model card,…
nvidia/Llama-3.1-Nemotron-Nano-4B-v1.1 · Hugging Face
NVIDIA has published Llama-3.1-Nemotron-Nano-4B-v1.1 on Hugging Face, a 4B-parameter reasoning and chat model derived from nvidia/Llama-3.1-Minitron-4B-Width-Base, which itself was built by compressing…
nvidia/AceReason-Nemotron-14B · Hugging Face
NVIDIA has published AceReason-Nemotron-14B, a 14B-parameter math and code reasoning model on Hugging Face. According to the model card, it was trained…
Replit makes OpenAI its default over Anthropic, CEO says | TechCrunch
Replit CEO and co-founder Amjad Masad said on stage at TechCrunch Disrupt that the AI coding platform no longer works with Anthropic…
Anthropic says its workforce has become less junior and is spending more time in meetings | TechCrunch
Anthropic published a report describing how its workforce composition and daily work have shifted over the past year, framing its own organization…
StepFun Releases Step 5 Preview Open Weights: 600B MoE, 27B Active, 1M Context - MarkTechPost
StepFun has released the weights for Step 5 Preview, a sparse Mixture-of-Experts model aimed at long-horizon agentic work, with FP8 and BF16…
20 Jev Alternatives: The Decision Models Taking On TypeSafe in 2026 - MarkTechPost
Per MarkTechPost, TypeSafe launched Jev on September 15, 2026 as a closed-weight decision model that scores predefined answer options (labels, scores, or…
DeepSeek Just Released V4.1-Flash-1017, Scores 77.0 on DeepSWE, Opens Search API - MarkTechPost
DeepSeek has released DeepSeek-V4.1-Flash-1017, an updated post-training checkpoint of its V4.1-Flash model, available through its website, app, and API under the deepseek-flash…
Meet Kyutai REKI: The 2.4B VLA Model Bringing Efficient Memory to Robot Manipulation - MarkTechPost
Source: MarkTechPost (Asif Razzaq), summarizing a release from Kyutai with Inria and École des Ponts. The source reports that Kyutai published REKI,…
MiniMax-M1 - a MiniMaxAI Collection
This Hugging Face collection page from MiniMaxAI lists the MiniMax-M1 model family, which the page describes as an open-weight, large-scale hybrid-attention reasoning…
katanemo/Arch-Agent-7B · Hugging Face
Katanemo has released Arch-Agent-7B, a 7.6-billion-parameter language model hosted on Hugging Face and aimed at function calling and agent workloads. According to…
google/magenta-realtime · Hugging Face
Google has released Magenta RealTime, an open-weights music generation model built from the same research as MusicFX DJ and Lyria RealTime. The…
GitHub - IBM/mcp-context-forge: An AI Gateway, registry, and proxy that sits in front of any MCP, A2A, or REST/gRPC APIs, exposing a unified endpoint with centralized discovery, guardrails and managem
IBM's open-source project mcp-context-forge, described in its GitHub README, is a gateway and registry that sits in front of Model Context Protocol…
Advancing Multimodal Reasoning: From Optimized Cold Start to Staged Reinforcement Learning
This arXiv paper (Chen et al., submitted June 2025, revised January 2026) studies how to train multimodal large language models (MLLMs) to…
HtFLlib: A Comprehensive Heterogeneous Federated Learning Library and Benchmark
Researchers including Jianqing Zhang and Qiang Yang describe HtFLlib, a Python-style library and benchmark for heterogeneous federated learning (HtFL), where clients train…
Small Language Models are the Future of Agentic AI
This arXiv position paper, authored by Peter Belcak and seven co-authors including NVIDIA-affiliated researchers, argues that small language models (SLMs) are sufficiently…
From Bytes to Ideas: Language Modeling with Autoregressive U-Nets
Researchers Mathurin Videau and five co-authors, posted to arXiv on 17 June 2025, propose a language model that does not rely on…
ProtoReasoning: Prototypes as the Foundation for Generalizable Reasoning in LLMs
A June 2025 arXiv preprint by Feng He and co-authors proposes that cross-domain generalization in Large Reasoning Models trained with long chain-of-thought…
Stream-Omni: Simultaneous Multimodal Interactions with Large Language-Vision-Speech Model
Researchers from the arXiv submission (authors Shaolei Zhang, Shoutao Guo, Qingkai Fang, Yan Zhou, and Yang Feng) describe Stream-Omni, a large language-vision-speech…
Go-Browse: Training Web Agents with Structured Exploration
Researchers Apurva Gandhi and Graham Neubig, as described in an arXiv preprint (first submitted June 2025, revised March 2026), present Go-Browse, a…
Reinforcement Learning Teachers of Test Time Scaling
This arXiv paper, by Edoardo Cetin, Tianyu Zhao, and Yujin Tang, proposes training a new class of reasoning language models called Reinforcement-Learned…
Agentic misalignment: How LLMs could be insider threats \ Anthropic
Anthropic reports results from stress-testing 16 leading models from multiple developers in simulated corporate environments where the models could autonomously send emails…
How we built our multi-agent research system \ Anthropic
Anthropic's engineering team describes the architecture and development of the multi-agent Research feature in Claude, published June 13, 2025. The system uses…
deepseek-ai/DeepSeek-R1-0528 · Hugging Face
DeepSeek has released DeepSeek-R1-0528, a minor version update to its R1 reasoning model, listed on Hugging Face. According to the model card,…
Perplexity Open-Sources SPACE, a Firecracker MicroVM Sandbox for AI Agents - MarkTechPost
Perplexity has open-sourced SPACE (Sandboxed Platform for Agentic Code Execution), the sandbox system it uses for every task in its Perplexity Computer…
GitHub - SkyworkAI/Skywork-Reward-V2: Scaling Preference Data Curation via Human-AI Synergy · GitHub
Skywork has released Skywork-Reward-V2, a series of eight Bradley-Terry reward models trained on a curated mixture of 26 million preference pairs. According…
Skywork-Reward-V2: Scaling Preference Data Curation via Human-AI Synergy
Per the arXiv paper (Liu et al., submitted July 2025, revised March 2026, ICLR 2026 poster), the authors argue that open reward…
osmosis-ai/Osmosis-Apply-1.7B · Hugging Face
Osmosis AI has published Osmosis-Apply-1.7B, a 1.7B-parameter model fine-tuned from Qwen3-1.7B to perform code merges, the step an AI code editor takes…
GitHub - bytedance/trae-agent: Trae Agent is an LLM-based agent for general purpose software engineering tasks. · GitHub
ByteDance has released Trae Agent, an open-source, LLM-based agent for general-purpose software engineering tasks, distributed as a command-line tool. The repository describes…
SmolLM3: smol, multilingual, long-context reasoner
Hugging Face's SmolLM3 is a fully open 3B-parameter language model, released with its complete training recipe, including architecture changes, data mixtures, and…
GitHub - googleapis/mcp-toolbox: MCP Toolbox for Databases is an open source MCP server for databases. · GitHub
This is the GitHub repository for MCP Toolbox for Databases, an open source Model Context Protocol (MCP) server that lets AI agents,…
GitHub - Google-Health/medsiglip · GitHub
This GitHub repository from Google-Health hosts MedSigLIP, a model based on SigLIP (Sigmoid Loss for Language Image Pre-training) trained to map medical…
GitHub - Google-Health/medgemma · GitHub
Google Health's GitHub repository for MedGemma describes a collection of Gemma 3 model variants adapted for medical text and image understanding, intended…
MedGemma Technical Report
This arXiv technical report (arXiv:2507.05201, first submitted 7 July 2025 and last revised 6 April 2026, per the listing) introduces MedGemma, a…
GTA1: GUI Test-time Scaling Agent
This arXiv preprint (submitted July 8, 2025; latest revision v5 in October 2025) by 15 authors introduces GTA1, a graphical user interface…
GitHub - QwenLM/qwen-code: An open-source AI coding agent that lives in your terminal. · GitHub
The QwenLM/qwen-code GitHub repository presents an open-source AI coding agent designed to run in a terminal, with additional interfaces including IDE plugins…
Kimi-K2 - a moonshotai Collection
This is a Hugging Face collection page from the Moonshot AI organization (moonshotai) grouping several of its open model releases under the…
microsoft/Phi-4-mini-flash-reasoning · Hugging Face
Microsoft has released Phi-4-mini-flash-reasoning, a 3.8B-parameter open model under the MIT license, aimed at multi-step mathematical reasoning in memory- or latency-constrained settings.…
Decoder-Hybrid-Decoder Architecture for Efficient Reasoning with Long Generation
Researchers from Microsoft and collaborators describe SambaY, a decoder-hybrid-decoder language model architecture built around a new mechanism called the Gated Memory Unit…
Audio Flamingo 3: Advancing Audio Intelligence with Fully Open Large Audio Language Models
Researchers from NVIDIA and collaborating institutions describe Audio Flamingo 3 (AF3), a large audio-language model that handles speech, sound, and music understanding…
nvidia/audio-flamingo-3 · Hugging Face
NVIDIA has released Audio Flamingo 3 (AF3), an open large audio-language model covering speech, environmental sounds, and music. According to the model…
OpenReasoning-Nemotron: A Family of State-of-the-Art Distilled Reasoning Models
NVIDIA announced OpenReasoning-Nemotron, a set of reasoning-focused LLMs distilled from DeepSeek R1 0528 (671B). Four sizes are released on Hugging Face at…
nvidia/OpenReasoning-Nemotron-1.5B · Hugging Face
NVIDIA has published OpenReasoning-Nemotron-1.5B on Hugging Face, a 1.5B-parameter reasoning model post-trained from Qwen2.5-1.5B for math, code, and science solution generation. The…
Build More Accurate and Efficient AI Agents with the New NVIDIA Llama Nemotron Super v1.5 | NVIDIA Technical Blog
NVIDIA announced Llama Nemotron Super v1.5, a 49B-parameter model released on a Friday in late July 2025, which the company says was…
nvidia/Llama-3_3-Nemotron-Super-49B-v1_5 · Hugging Face
NVIDIA has released Llama-3.3-Nemotron-Super-49B-v1.5, a 49B-parameter reasoning and chat model derived from Meta's Llama-3.3-70B-Instruct. According to the model card on Hugging Face,…
Qwen/Qwen3-Coder-480B-A35B-Instruct · Hugging Face
Qwen has published Qwen3-Coder-480B-A35B-Instruct on Hugging Face, the largest variant in its Qwen3-Coder family, which the model card describes as the team's…
zai-org/GLM-4.5-Air · Hugging Face
Z.ai (zai-org) published GLM-4.5-Air, a compact member of its GLM-4.5 series of foundation models aimed at intelligent-agent use. The model has 106…
zai-org/GLM-4.5 · Hugging Face
Z.ai (zai-org) published GLM-4.5, a family of open-weight foundation models positioned for agentic applications, combining reasoning, coding, and tool use. The flagship…
Ricursive Intelligence is betting on AI-designed chips to help solve the compute crisis | TechCrunch
Per TechCrunch reporter Aisha Malik, Ricursive Intelligence co-founders Anna Goldie and Azalia Mirhoseini presented at TechCrunch Disrupt on their startup's work on…
GitHub - tsinghua-fib-lab/AgentSociety: AgentSociety 2 is a modern, LLM-native agent simulation platform designed for social science research and experimental design. It provides a flexible framework
The source is the GitHub repository for AgentSociety, a framework from Tsinghua's FIB Lab for running LLM-based agent simulations. The repository describes…
Adaptively Robust LLM Inference Optimization under Prediction Uncertainty
Researchers Zixi Chen, Yinyu Ye, and Zijie Zhou (arXiv:2508.14544, first submitted 20 August 2025, revised 1 September 2025) study how to schedule…
GitHub - next-state/open-dreamer: Open-source Dreamer world-model implementation in JAX · GitHub
The next-state/open-dreamer repository is an open-source implementation of the Dreamer 4 world-model pipeline written in JAX/Flax NNX. It is trained on Minecraft/VPT-style…
Falcon-H1: A Family of Hybrid-Head Language Models Redefining Efficiency and Performance
This arXiv technical report (submitted 30 July 2025, 27 authors) introduces Falcon-H1, a family of large language models that combine Transformer attention…
Galileo: Learning Global & Local Features of Many Remote Sensing Modalities
Per the arXiv listing (arXiv:2502.09356, submitted 13 Feb 2025, revised through v3 on 4 Jun 2025), the authors present Galileo, a highly…
GitHub - google/langextract: A Python library for extracting structured information from unstructured text using LLMs with precise source grounding and interactive visualization. · GitHub
Google's langextract repository, as shown on its GitHub page, describes a Python library that uses large language models to pull structured data…
zai-org/GLM-4.5V · Hugging Face
Source: Hugging Face model card for zai-org/GLM-4.5V (https://huggingface.co/zai-org/GLM-4.5V). The card describes GLM-4.5V, a vision-language model in the GLM-V family from ZhipuAI, built…
DINOv3 - a facebook Collection
This is a Hugging Face collection page from Meta (listed under the 'facebook' organization) grouping releases of DINOv3, a family of self-supervised…
microsoft/VibeVoice-1.5B · Hugging Face
Microsoft Research has published VibeVoice-1.5B, an open-weight text-to-speech model on Hugging Face aimed at long-form, multi-speaker conversational audio such as podcasts. According…
nvidia/parakeet-tdt-0.6b-v3 · Hugging Face
This is the Hugging Face model card for NVIDIA's parakeet-tdt-0.6b-v3, a multilingual automatic speech recognition (ASR) model with roughly 0.6 billion parameters.…
nvidia/canary-1b-v2 · Hugging Face
This Hugging Face model card describes NVIDIA's canary-1b-v2, a 1-billion-parameter multilingual speech model that is listed under automatic speech recognition, with the…
nvidia/Granary · Datasets at Hugging Face
The page is the Hugging Face dataset card for nvidia/Granary, a multilingual speech dataset published by NVIDIA. The dataset is organized into…
GitHub - Accenture/mcp-bench: MCP-Bench: Benchmarking Tool-Using LLM Agents with Complex Real-World Tasks via MCP Servers · GitHub
Accenture has released MCP-Bench, an open-source framework for evaluating how well large language models use tools via the Model Context Protocol (MCP).…
google/embeddinggemma-300m · Hugging Face
Google DeepMind has released EmbeddingGemma, a 300M-parameter open text embedding model built from Gemma 3 with T5Gemma initialization, according to its Hugging…
TildeAI/TildeOpen-30b · Hugging Face
Tilde.ai has published TildeOpen-30b, a 30-billion-parameter dense decoder-only transformer on Hugging Face under a CC-BY-4.0 license. The model is positioned as a…
baidu/ERNIE-4.5-21B-A3B-Thinking · Hugging Face
This Hugging Face model card, published by Baidu, describes ERNIE-4.5-21B-A3B-Thinking, a post-trained text Mixture-of-Experts model. According to the card, the model has…
Jet-Nemotron: Efficient Language Model with Post Neural Architecture Search
A tech report on arXiv (submitted 21 Aug 2025, revised through v3 on 28 Sep 2025) describes Jet-Nemotron, a family of hybrid-architecture…
nvidia/canary-1b · Hugging Face
NVIDIA's model card on Hugging Face describes Canary-1B, a 1-billion-parameter multilingual speech model from the NeMo family. It supports automatic speech recognition…
nvidia/parakeet-rnnt-1.1b · Hugging Face
This is the Hugging Face model card for parakeet-rnnt-1.1b, an automatic speech recognition model that transcribes lowercase English text. The model is…
Introducing Claude Sonnet 5 \ Anthropic
Anthropic has released Claude Sonnet 5, which it describes as its most agentic Sonnet-class model to date, capable of planning, using tools…
On the Theoretical Limitations of Embedding-Based Retrieval
This arXiv paper (accepted to ICLR'26, first submitted August 2025, revised March 2026) argues that single-vector embedding models face a fundamental capacity…
REFRAG: Rethinking RAG based Decoding
A research paper posted to arXiv (v1 September 1, 2025; v2 October 12, 2025) by Xiaoqiang Lin, Aritra Ghosh, Bryan Kian Hsiang…
Scaling Recommender Transformers to One Billion Parameters
The paper, posted to arXiv by Kirill Khrylchenko, Artem Matveev, Sergei Makeev and Vladimir Baikalov (cs.IR) and submitted to KDD 2026, presents…
ParaThinker: Native Parallel Thinking as a New Paradigm to Scale LLM Test-time Compute
Researchers Hao Wen and six co-authors posted an arXiv preprint (2509.04475, submitted 30 August 2025) proposing ParaThinker, an end-to-end framework that trains…
RL's Razor: Why Online Reinforcement Learning Forgets Less
Per an arXiv preprint submitted 4 September 2025 by Idan Shenfeld, Jyothish Pari, and Pulkit Agrawal, fine-tuning with on-policy reinforcement learning (RL)…
Granite Docling - a ibm-granite Collection
This is a Hugging Face collection page from IBM's ibm-granite organization, dedicated to Granite Docling. According to the page, the models are…
google/vaultgemma-1b · Hugging Face
Google has released VaultGemma, a 1-billion-parameter open language model pre-trained from scratch with differential privacy, per its Hugging Face model card. Training…
facebook/MobileLLM-R1-950M · Hugging Face
This Hugging Face model card describes facebook/MobileLLM-R1-950M, a text-generation model published by Meta with the transformers and safetensors libraries. The repository reports…
GitHub - bentoml/llm-optimizer: Benchmark and optimize LLM inference across frameworks with ease · GitHub
The BentoML team's llm-optimizer is an open-source Python tool for benchmarking and tuning inference performance of open-source LLMs. It runs benchmarks against…
ibm-granite/granite-embedding-small-english-r2 · Hugging Face
IBM's Granite Embedding team has published granite-embedding-small-english-r2 on Hugging Face, a 47M-parameter dense bi-encoder for English text embeddings. It outputs 384-dimensional vectors,…
ibm-granite/granite-embedding-english-r2 · Hugging Face
IBM's Granite Embedding team released granite-embedding-english-r2 on August 15, 2025, a 149M-parameter dense bi-encoder that produces 768-dimensional embeddings for English text with…
GitHub - MoonshotAI/checkpoint-engine: Checkpoint-engine is a simple middleware to update model weights in LLM inference engines · GitHub
MoonshotAI has released checkpoint-engine, an open-source middleware library that updates model weights in place inside LLM inference engines. The project targets reinforcement…
GitHub - google-agentic-commerce/AP2: Building a Secure and Interoperable Future for AI-Driven Payments. · GitHub
The source is the GitHub repository for the Agent Payments Protocol (AP2), an open project under the google-agentic-commerce organization that provides code…
GitHub - Alibaba-NLP/DeepResearch: Tongyi Deep Research, the Leading Open-source Deep Research Agent · GitHub
Alibaba's Tongyi Lab has published Tongyi DeepResearch, an agentic large language model with 30.5 billion total parameters, of which about 3.3 billion…
Alibaba-NLP/Tongyi-DeepResearch-30B-A3B · Hugging Face
Alibaba's Tongyi Lab has published Tongyi DeepResearch, an open agentic large language model with 30 billion total parameters, of which about 3…
Qwen/Qwen3-Next-80B-A3B-Thinking-FP8 · Hugging Face
Alibaba's Qwen team published an FP8-quantized checkpoint of Qwen3-Next-80B-A3B-Thinking on Hugging Face, the thinking-mode variant of its Qwen3-Next series. The model card…
Qwen/Qwen3-Next-80B-A3B-Instruct-FP8 · Hugging Face
Qwen has released Qwen3-Next-80B-A3B-Instruct-FP8, an FP8-quantized version of its Qwen3-Next instruct model, according to the Hugging Face model card. The model has…
VoXtream: Full-Stream Text-to-Speech with Extremely Low Latency
Researchers Nikita Torgashov, Gustav Eje Henter, and Gabriel Skantze describe VoXtream, a zero-shot streaming text-to-speech system designed for real-time use, as reported…
GitHub - facebookresearch/cwm: Research code artifacts for Code World Model (CWM) including inference tools, reproducibility, and documentation. · GitHub
Meta's FAIR CodeGen team has published the research artifacts for Code World Model (CWM), a 32-billion-parameter open-weights LLM aimed at code generation…
Gemini Robotics 1.5 brings AI agents into the physical world — Google DeepMind
Google DeepMind has introduced two robotics models built on the Gemini family. Gemini Robotics-ER 1.5 is an embodied reasoning vision-language model that…
smol2operator Release - a smolagents Collection
This Hugging Face collection, maintained under the smolagents organization, bundles the release artifacts for smol2operator, a GUI agent project. The collection was…
Smol2Operator: Post-Training GUI Agents for Computer Use
Hugging Face published a blog post describing a two-phase supervised fine-tuning pipeline that gives the lightweight vision-language model SmolVLM2-2.2B-Instruct GUI grounding and…
GitHub - deepseek-ai/DeepSeek-V3.2-Exp · GitHub
DeepSeek has released DeepSeek-V3.2-Exp, an experimental model positioned as an intermediate step toward its next-generation architecture. Built on V3.1-Terminus, it adds DeepSeek…
GitHub - Mega4alik/ollm · GitHub
The repository for oLLM, an open-source Python library built on Hugging Face Transformers and PyTorch, is described as enabling large-context offline LLM…
GitHub - QwenLM/Qwen3Guard: Qwen3Guard is a multilingual guardrail model series developed by the Qwen team at Alibaba Cloud. · GitHub
The Qwen team at Alibaba Cloud has published Qwen3Guard, a series of safety moderation models built on Qwen3 and trained on a…
GitHub - SakanaAI/ShinkaEvolve: ShinkaEvolve: Towards Open-Ended and Sample-Efficient Program Evolution 🧬 · GitHub
SakanaAI's ShinkaEvolve is an open-source Python framework, described in its GitHub repository, that combines large language models with evolutionary search to optimize…
Here are the top AI agents that can live in your text messages | TechCrunch
TechCrunch's Lauren Forristal surveys a growing category of AI agents that users interact with through SMS, iMessage, WhatsApp or Telegram rather than…
Apple discloses deal to hire team and license tech from personalized podcast startup Huxe | TechCrunch
Per a regulatory filing that Apple made with the European Commission, the company has agreed to offer employment to certain employees of…
Nace AI Open-Sources Drex 1.5: A 9B Decision Model That Scores Options, Not Text - MarkTechPost
Nace.AI has released Drex 1.5, an open-weights 8.95B-parameter dense decision model that does not generate text. Instead, it takes a state (text…
Microsoft AI Releases Microsoft-Decision-1: A Qwen3.5-9B Decision-Scoring Model - MarkTechPost
Microsoft has released Microsoft-Decision-1, a model that scores a fixed set of answer options and returns a calibrated probability for each, rather…
Sakana AI’s LLM Peer Review System Catches 73% of Core-Claim Errors - MarkTechPost
MarkTechPost, summarizing a Sakana AI paper titled "Beyond Imitation" published in TMLR, reports on an LLM-assisted peer review approach evaluated by whether…
ShinkaEvolve: Towards Open-Ended And Sample-Efficient Program Evolution
Researchers Robert Tjarko Lange, Yuki Imajuku, and Edoardo Cetin posted arXiv preprint 2509.19349 on 17 September 2025, introducing ShinkaEvolve, an open-source framework…
Granite 4.0 Language Models - a ibm-granite Collection
This is a Hugging Face collection page from IBM's ibm-granite organization listing the Granite 4.0 family of language models. The collection is…
ServiceNow-AI/Apriel-1.5-15b-Thinker · Hugging Face
ServiceNow's Apriel-1.5-15b-Thinker is a 15-billion-parameter multimodal reasoning model released on Hugging Face under the MIT license, the second model in the company's…
LiquidAI/LFM2-Audio-1.5B · Hugging Face
Liquid AI has published LFM2-Audio-1.5B on Hugging Face, described by the model card as the company's first end-to-end audio foundation model. The…
zai-org/GLM-4.6 · Hugging Face
Z.ai (zai-org) published GLM-4.6 on Hugging Face as a text-generation model in the transformers/safetensors format, released under an MIT license. According to…
Introducing CodeMender: an AI agent for code security — Google DeepMind
Google DeepMind has shared early research results on CodeMender, an AI agent designed to improve software security automatically. According to the post,…
Less is More: Recursive Reasoning with Tiny Networks
This arXiv preprint (submitted 6 October 2025 by Alexia Jolicoeur-Martineau) proposes the Tiny Recursive Model (TRM), a recursive reasoning approach presented as…
GitHub - meridianlabs-ai/inspect_petri: An alignment auditing agent capable of quickly exploring alignment hypothesis · GitHub
This GitHub repository, published under the meridianlabs-ai organization and tied to the safety-research/petri project, hosts Inspect Petri, an open-source auditing agent for…
GitHub - google-gemini/computer-use-preview · GitHub
Source: GitHub repository google-gemini/computer-use-preview (github.com/google/computer-use-preview). The repository is a Python command-line agent that uses Gemini models to operate a browser on a…
LIMI: Less is More for Agency
A paper posted to arXiv (submitted 22 September 2025, revised 25 September 2025) by a 21-author team introduces LIMI, a method for…
Regression Language Models for Code
This arXiv paper (Akhauri, Song, Wongpanich, Lewandowski, and Abdelfattah) studies code-to-metric regression, the task of predicting numeric outcomes of code execution directly…
TUMIX: Multi-Agent Test-Time Scaling with Tool-Use Mixture
Source: arXiv preprint 2510.01279 (submitted 30 Sep 2025 by Yongchao Chen and eight co-authors, cs.CL/cs.AI). The paper presents Tool-Use Mixture (TUMIX), an…
Salesforce/CoDA-v0-Instruct · Hugging Face
Salesforce AI Research has released CoDA-1.7B-Instruct, a 1.7-billion-parameter diffusion-based language model aimed at code generation and completion, along with a CoDA-1.7B-Base foundation…
GitHub - workersio/skills: Agent skills to find and fix software bugs · GitHub
The workersio/skills GitHub repository publishes a packaged agent skill, named @workersio/skills, aimed at coding agents such as Codex and Claude Code. Its…
poolside/Laguna-XS-2.1 · Hugging Face
Poolside has released Laguna XS 2.1, a 33B-total-parameter Mixture-of-Experts model with 3B parameters activated per token, aimed at agentic coding and long-horizon…
fdtn-ai/antares-350m · Hugging Face
Cisco's Foundation AI group has released Antares-350M, a 350-million-parameter open-weight language model built on IBM Granite 4.0 350M and specialized for vulnerability…
nvidia/Cosmos3-Edge · Hugging Face
NVIDIA has published Cosmos3-Edge, a 4-billion-parameter member of its Cosmos3 family of omnimodal world models, per the Hugging Face model card. The…
Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models
Researchers introduced ACE (Agentic Context Engineering), a framework that improves language model performance by treating contexts as evolving playbooks rather than static…
Agent Learning via Early Experience
Researchers present a training paradigm called "early experience" for language agents that leverages the agent's own interactions rather than relying solely on…
GitHub - OpenTSLM/OpenTSLM: OpenTSLM: Time-Series Language Models for Reasoning over Multivariate Medical Text- and Time-Series Data · GitHub
OpenTSLM is an open-source framework that extends large language models to natively handle time-series data for medical applications. Developed by Stanford University…
LiquidAI/LFM2-8B-A1B · Hugging Face
Liquid AI has released LFM2-8B-A1B, a hybrid mixture-of-experts language model with 8.3 billion total parameters and 1.5 billion active parameters, designed for…
inclusionAI/Ring-1T · Hugging Face
inclusionAI released Ring-1T, an open-source trillion-parameter mixture-of-experts language model featuring 1 trillion total parameters with 50 billion activated parameters and support for…
Weak-for-Strong: Training Weak Meta-Agent to Harness Strong Executors
Researchers propose Weak-for-Strong Harnessing (W4S), a framework that trains smaller language models to automatically design and optimize workflows for deploying stronger models…
Qwen3-VL - a Qwen Collection
Alibaba's Qwen has released the Qwen3-VL collection, a series of vision-language models ranging from 2B to 236B parameters. The collection includes multiple…
GitHub - NVlabs/QeRL: [ICLR 2026]QeRL enables RL for 32B LLMs on a single H100 GPU. · GitHub
QeRL is a reinforcement learning framework from NVIDIA Labs that combines NVFP4 quantization with Low-Rank Adaptation (LoRA) to reduce memory and computational…
deepseek-ai/DeepSeek-OCR · Hugging Face
DeepSeek released DeepSeek-OCR, a vision-language model specialized for optical character recognition and document processing, hosted on Hugging Face. The model uses a…
BitNet Distillation
Researchers from Microsoft and affiliated institutions published BitNet Distillation (BitDistill), a technique for converting full-precision language models into 1.58-bit ternary weights (values…
Anthropic Releases Claude Haiku 5.5: A Small Model With 1M Context Priced at $0.10 per Million Input Tokens - MarkTechPost
Anthropic released Claude Haiku 5.5, a small language model priced at $0.10 per million input tokens and $0.50 per million output tokens…
VISTA: A Test-Time Self-Improving Video Generation Agent
Researchers introduced VISTA, a multi-agent framework that autonomously improves text-to-video generation through iterative prompt refinement. The system decomposes user requests into structured…
GitHub - SalesforceAIResearch/WALT: Code for WALT – Web Agents that Learn Tools · GitHub
Salesforce AI Research released WALT (Web Agents that Learn Tools), an open-source framework enabling LLM agents to automatically discover and generate reusable…
GitHub - ovg-project/kvcached: Virtualized Elastic KV Cache for Dynamic GPU Sharing and Beyond · GitHub
kvcached is an open-source KV cache library that enables elastic memory allocation for LLM serving and training on shared GPUs. It decouples…
Glyph: Scaling Context Windows via Visual-Text Compression
Researchers propose Glyph, a framework that addresses the computational and memory costs of scaling LLM context windows to million-token levels by rendering…
zai-org/Glyph · Hugging Face
Glyph is a 9-billion-parameter vision-language model released on Hugging Face (zai-org/Glyph) that addresses long-context processing by rendering text into images and processing…
GitHub - thu-coai/Glyph: Official Repository for "Glyph: Scaling Context Windows via Visual-Text Compression" · GitHub
Glyph is a framework developed by Tsinghua University's COAI group for extending context windows by rendering long text sequences into images and…
GitHub - microsoft/agent-lightning: The absolute trainer to light up AI agents. · GitHub
Microsoft released Agent Lightning v1.0, an open-source reinforcement learning framework designed to train AI agents without modifying existing agent harnesses or codebases.…