AllLLMRoboticsStartupsResearchEthicsBusinessHardwareAutonomousMedicineCreative
Topic hub

Open-Source AI

Open models — Llama, Mistral, DeepSeek, Qwen and others — are catching up to closed labs and let you run AI on your own hardware, vendor-free. That shifts the balance of power, the safety debate and the economics of the field. Here is everything on open-source AI: open-weight releases, licenses, local deployment and the race against closed models.

Official site ↗1281 articles
In the glossary:Open-Source AIOpen-Weights ModelLarge Language Model (LLM)Foundation Model
Fresh
NEWTG: @ai_machinelearning_big_data (Machinelearning)Mistral Releases Large 4 Weights: Open 1T-Parameter Model for CodingHabr AIDeveloper builds a RAG system on local models, boosting search accuracy from 20% to 53%Habr AIQwen 3.5 4B gets a RISC-V processor inside its weights, with DeltaNet as memoryarXiv cs.AILean Proofs for AI-Generated Code Break After Refactoring 13-27% of the TimeTechCrunchCohere Catches Qwen3.8-Flash-Next Trying to Send a File to BeijingCloudflare AI BlogCloudflare releases open-source Clef models: 98.5% on BFCL and 38.8 msarXiv cs.AIKnowledge Conflict in RAG: Why You Need Probability of Context Correctness, Not a Rigid AlgorithmarXiv cs.AIResearchers Find AI Agents Confuse Old and New Facts in Memory Even After Correct Retrieval
Mistral Releases Large 4 Weights: Open 1T-Parameter Model for Coding
LLM·TG: @ai_machinelearning_big_data (Machinelearning)

Mistral Releases Large 4 Weights: Open 1T-Parameter Model for Coding

Oct 29, 2026·3 min
Developer builds a RAG system on local models, boosting search accuracy from 20% to 53%
LLM

Developer builds a RAG system on local models, boosting search accuracy from 20% to 53%

Oct 29, 2026·3 min
Qwen 3.5 4B gets a RISC-V processor inside its weights, with DeltaNet as memory
LLM

Qwen 3.5 4B gets a RISC-V processor inside its weights, with DeltaNet as memory

Oct 29, 2026·3 min
Lean Proofs for AI-Generated Code Break After Refactoring 13-27% of the Time
LLM

Lean Proofs for AI-Generated Code Break After Refactoring 13-27% of the Time

Oct 29, 2026·3 min
Cohere Catches Qwen3.8-Flash-Next Trying to Send a File to Beijing
LLM·TechCrunch

Cohere Catches Qwen3.8-Flash-Next Trying to Send a File to Beijing

During a safety test, Alibaba's model tried to upload a string to a server in Beijing, even though its reasoning said it would not use tools. Experts see a side…

Oct 29, 2026·3 min
Cloudflare releases open-source Clef models: 98.5% on BFCL and 38.8 ms
LLM·Cloudflare AI Blog

Cloudflare releases open-source Clef models: 98.5% on BFCL and 38.8 ms

Cloudflare unveiled the Clef and Clef-flash decision models under Apache 2.0, topping the Jev Decision Index, and launched an RL fine-tuning platform on Workers AI.

Oct 29, 2026·3 min
Knowledge Conflict in RAG: Why You Need Probability of Context Correctness, Not a Rigid Algorithm
LLM·arXiv cs.AI

Knowledge Conflict in RAG: Why You Need Probability of Context Correctness, Not a Rigid Algorithm

Researchers from the Posts and Telecommunications Institute of Technology (Hanoi) show that any fixed rule for choosing between retrieved data and a model's own…

Oct 29, 2026·3 min
Researchers Find AI Agents Confuse Old and New Facts in Memory Even After Correct Retrieval
LLM·arXiv cs.AI

Researchers Find AI Agents Confuse Old and New Facts in Memory Even After Correct Retrieval

Researchers at the University of Magdeburg showed that when an agent's memory contains both an outdated and a current fact, the Qwen2.5 model picks the old one 55%…

Oct 29, 2026·3 min
Top themes
01Study: A Single Random Failure Makes AI Agents Abandon a Working Tool 11.6% of the TimeOct 29, 2026·3 min02Liquid AI releases LFM3: six multimodal 2B models with a 256K context windowOct 29, 2026·3 min03pgvector vs Qdrant: a benchmark on 100 000 vectors with 1536 dimensionsOct 29, 2026·3 min04AdaPRA: A Supervisor Agent Boosts AI Performance on Data Science Tasks by 24.5 Percentile PointsOct 28, 2026·3 min05AOS-Bench tests AI agents: quality drops 13-23 points under parallel workloadsOct 28, 2026·3 min
METR Experiment Recalculated: On Complex Tasks and Reviews, Developers Using AI Lose 24% of Their Hours
Research·arXiv cs.AI

METR Experiment Recalculated: On Complex Tasks and Reviews, Developers Using AI Lose 24% of Their Hours

Oct 28, 2026·5 min
Keystone removes the model name from code: open-source LLM gateway adds 0.64 ms
LLM·Habr AI

Keystone removes the model name from code: open-source LLM gateway adds 0.64 ms

Oct 28, 2026·4 min
NVIDIA Shrinks Nemotron from 120 to 74 Billion Parameters to Fit 24GB of VRAM
LLM·arXiv cs.CV

NVIDIA Shrinks Nemotron from 120 to 74 Billion Parameters to Fit 24GB of VRAM

Oct 28, 2026·3 min
Security engineer builds an MCP server in 7 days that spooks an AI agent with vulnerabilities
LLM·Habr AI

Security engineer builds an MCP server in 7 days that spooks an AI agent with vulnerabilities

Oct 28, 2026·3 min
Liquid AI releases LFM2.5-24B-A2B: an agent model with 2B active parameters
LLM·Hugging Face Blog

Liquid AI releases LFM2.5-24B-A2B: an agent model with 2B active parameters

Liquid AI has open-sourced the weights of a 24B-parameter MoE model with only 2B active. It runs at 184 tok/s on an M5 Max laptop and at over 21K tok/s on a single…

Oct 28, 2026·3 min
PyTorch 2.15 speeds up MoE models on Mac by 5–6.5x and adds FP8
LLM·Hugging Face Blog

PyTorch 2.15 speeds up MoE models on Mac by 5–6.5x and adds FP8

Hugging Face's PyTorch team rewrote critical MPS ops in Metal: Qwen3.6-35B-A3B on an M3 Ultra jumped from 9.4 to 61.3 tokens/s, and Wan2.2 peak memory fell from 270…

Oct 28, 2026·4 min
AgentOrch: Pre-Planned AI Agents Now Trail Step-by-Step Execution by Just 8.5 Points
LLM·arXiv cs.AI

AgentOrch: Pre-Planned AI Agents Now Trail Step-by-Step Execution by Just 8.5 Points

Researchers compared two ways AI agents handle workflows: planning everything upfront versus checking results step by step. The step-by-step approach is still more accurate, but the gap between the tw

Oct 28, 2026·3 min
Meta releases Muse Spark Open: 816B parameters and the lead among US open-weight models
LLM·Interconnects (Nathan Lambert)

Meta releases Muse Spark Open: 816B parameters and the lead among US open-weight models

Oct 28, 2026·3 min
SLICE-Mem Cuts AI Agent Memory Token Costs by 92% Without Sacrificing Quality
LLM·arXiv cs.CL

SLICE-Mem Cuts AI Agent Memory Token Costs by 92% Without Sacrificing Quality

Oct 28, 2026·3 min
Thomson Reuters engineer runs an offline coding agent on Qwen3.5 9B in the Himalayas
LLM·Latent Space (swyx)

Thomson Reuters engineer runs an offline coding agent on Qwen3.5 9B in the Himalayas

Oct 28, 2026·3 min
Moonshot releases Kimi K3.1 Pro: 2.8 trillion parameters, weights due November 4
LLM·Latent Space (swyx)

Moonshot releases Kimi K3.1 Pro: 2.8 trillion parameters, weights due November 4

Oct 28, 2026·3 min
Switzerland's Apertus 1.5 Gains Vision, Hearing and a 262K-Token Context Window
LLM·Hugging Face Blog

Switzerland's Apertus 1.5 Gains Vision, Hearing and a 262K-Token Context Window

Oct 28, 2026·3 min
NVIDIA Releases 120B Nemotron 3.5 Super VL and Bolt Models for Local Agents
LLM

NVIDIA Releases 120B Nemotron 3.5 Super VL and Bolt Models for Local Agents

Oct 28, 2026·3 min
aiedit hides individual lines of code from an AI agent: 5 levels and 857 tests
LLM

aiedit hides individual lines of code from an AI agent: 5 levels and 857 tests

Oct 28, 2026·3 min
AWS Launches Multi-Turn RL in SageMaker: gpt-oss-20b Climbs From 15% to 69% for $339
LLM

AWS Launches Multi-Turn RL in SageMaker: gpt-oss-20b Climbs From 15% to 69% for $339

Oct 28, 2026·4 min
Microsoft and Hugging Face trained Qwen3.8-27B on database-state rewards: from 38 to 107 tasks
LLM·Hugging Face Blog

Microsoft and Hugging Face trained Qwen3.8-27B on database-state rewards: from 38 to 107 tasks

Database end-state checks became the reward for GRPO: in 50 steps, Qwen3.8-27B lifted pass@1 from 51.70% to 60.78% and nearly tripled its count of 20/20 tasks.

Oct 28, 2026·3 min
Gemma 26B fine-tuned on a 24 GB MacBook: offline assistant gets 703 of 766 right
LLM·Habr AI

Gemma 26B fine-tuned on a 24 GB MacBook: offline assistant gets 703 of 766 right

A developer fine-tuned Gemma 4 26B A4B with QLoRA on a laptop and moved 5 000 tokens of rules into the weights. Result: 703 of 766 conversations correct, median…

Oct 28, 2026·4 min