OsAi

Trying to keep up with the open source AI revolution in 2025 !

This repository serves as a logbook to document the ongoing Open Source AI (OsAI) revolution. Currently, it is a compilation of links that aim to capture the daily unfolding of this event. Occasionally, it also serves as an entry point to explore more fundamental concepts. This is truly an open work in progress. If you wish to discuss this, or contribute, please fork this repository and submit a Pull Request. Thank you!

NEW! : Heatmap and models

Previously

• Pre-R1

January

• Jan. 20 : Release DeepSeek-R1 🚩
• Jan. 20 : We are living in a timeline where a non-US company is keeping the original mission of OpenAI alive (LinkedIn)
• Jan. 21 : Stargate LLC announced by Donald Trump
• Jan. 21 : Kimi-k1.5
• Jan. 22 : DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
• Jan. 24 : Open-r1 : Fully open reproduction of DeepSeek-R1
• Jan. 26 : AI leaders clash over safety and 100ドルbn Stargate project
• Jan. 27 : Nvidia loses nearly 600ドル billion in market
• Jan. 27 : DeepSeek Is A Chinese Security Nightmare Come True
• Jan. 28 : Open-R1: a fully open reproduction of DeepSeek-R1
• Jan. 28 : ICYMI: DeepSeek Is an Open Source Success Story
• Jan. 29 : DeepSeek R1 is now available on Azure
• Jan. 29 : On DeepSeek and Export Controls
• Jan. 29 : Open Source Impact on AI (LinkedIn/Bloomberg)
• Jan. 29 : International AI Safety Report (PDF)
• Jan. 31 : The DeepSeek Narrative Takes the World by Storm
• Jan. 31 : Some notes on "DeepSeek and export control"
• Jan. 31 : AMA with OpenAI's Sam Altman 'we have been on the wrong side of history'

February

• Feb. 1 : An analysis of DeepSeek's R1-Zero and R1
• Feb. 1 : Searching for DeepSeek's glitch tokens
• Feb. 1 : The Illustrated DeepSeek-R1
• Feb. 2 : Introducing deep research
• Feb. 2 : Open-R1 progress update in 13 points
• Feb. 3 : DeepSeek gives Europe's tech firms a chance to catch up
• Feb. 3 : DeepSeek, China & AI
• Feb. 4 : Open Deep Research
• Feb. 5 : DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models
• Feb. 7 : DeepSeek, data privacy on lawmakers' radar
• Feb. 9 : Trump’s AI ambition and China’s DeepSeek overshadow an AI summit in Paris
• Feb. 9 : Macron announces 'French Stargate' with funding from MGX
• Feb. 10 : A Deep Peek at DeepSeek
• Feb. 11 : Open Source AI's Breakneck Innovation
• Feb. 11 : US's Vance warns Europe on AI overreach, slams China's tech crackdown at Paris summit
• Feb. 12 : Eric Schmidt warns west to focus on open-source AI in competition with China
• Feb. 12 : China opposes 'politicising' technology after JD Vance's AI warning
• Feb. 12 : China's DeepSeek Is Changing the AI Race—What is Europe's Edge?
• Feb. 14 : China's Baidu to make latest Ernie AI model open-source as competition heats up
• Feb. 16 : DeepSeek drives 1ドル.3 trillion China stock rally as funds pile in
• Feb. 17 : Security Researchers Issue Stark Warning: Do Not Use DeepSeek-R1
• Feb. 17 : South Korea pauses downloads of DeepSeek AI over privacy concerns
• Feb. 17 : 'Game on': Tech execs say DeepSeek ramps up China-U.S. competition but won’t hurt OpenAI
• Feb. 18 : Open-Reasoner-Zero
• Feb. 21 : DeepSeek Open Infra
• Feb. 22 : Fast GRPO Fine-Tuning in 50 Minutes
• Feb. 24 : Microsoft Has Dropped Some AI Data Center Leases

March

• Mar. 2 : AI companies race to use 'distillation' to produce cheaper models
• Mar. 4 : Light-R1: Surpassing R1-Distill in Math from Scratch* with 1000ドル
• Mar. 5 : Alibaba QwQ-32B
• Mar. 5 : Leave it to Manus
• Mar. 6 : Alibaba's New QwQ 32B Model is as Good as DeepSeek-R1
• Mar. 6 : An open source collective for inter-agent collaboration
• Mar. 7 : Was Manus another DeepSeek moment?
• Mar. 7 : Chinese AI agent runs 50 social media accounts 24/7 automatically
• Mar. 9 : Manus probably isn't China's second 'DeepSeek moment'
• Mar. 11 : Open R1: Update #3
• Mar. 12 : Search-R1: Leverage Search Engines with Reinforcement Learning
• Mar. 13 : AI and Trust: Leading the Charge in an Era of Accelerated Innovation
• Mar. 19 : AI Policy: Response to the White House
• Mar. 20 : A Hugging Face Project Is Uncovering DeepSeek-R1's Secrets
• Mar. 21 : Restricting the use of DeepSeek at UBC
• Mar. 24 : DeepSeek-V3-0324
• Mar. 24 : China's open-source embrace upends conventional wisdom around artificial intelligence
• Mar. 25 : DeepSeek narrows China-US AI gap to three months, 01.AI founder Lee Kai-fu says
• Mar. 25 : Chinese AI startups change business models after DeepSeek's boom
• Mar. 31 : Sam Altman : 'we are planning to release our first open-weigh language model since GPT-2'

April

• Apr. 1 : DeepMind slows down research releases in battle to keep competitive edge
• Apr. 3 : DeepSeek : Inference-Time Scaling for Generalist Reward Modeling
• Apr. 5 : The Llama 4 herd: The beginning of a new era of natively multimodal AI innovation
• Apr. 7 : Llama-3.1-Nemotron-Ultra-253B-v1 by Nvidia
• Apr. 7 : Meta got caught gaming AI benchmarks
• Apr. 8 : Cogito v1 Preview
• Apr. 8 : DeepCoder: A Fully Open-Source 14B Coder at O3-mini Level
• Apr. 9 : Agent2Agent Protocol (A2A)
• Apr. 10 : Seed-Thinking-v1.5
• Apr. 16 : Granite-3.3-2B-Instruct
• Apr. 16 : d1: Diffusion Large Language Models
• Apr. 25 : Pleias-RAG
• Apr. 28 : Mem0: Building Production-Ready AI Agents
• Apr. 29 : Qwen3
• Apr. 30 : DeepSeek-Prover-V2
• Apr. 30 : Phi-4-reasoning-plus
• Apr. 30 : Helium 1
• Apr. 30 : Open source has a 'massive role to play' in AI orchestration platforms

May

• May 6 : Learning 'Transformer Reinforcement Learning'
• May 7 : Empowering multi-agent apps with the open Agent2Agent (A2A) protocol
• May 8 : Sam Altman testifies on AI competition before Senate committee
• May 8 : « It is a race that no company or country can win by itself »
• May 11 : OpenR1 Mixture-of-Thoughts
• May 14 : Insights into DeepSeek-V3
• May 19 : Introducing NLWeb
• May 19 : VS Code: Open Source AI Editor
• May 19 : MMaDA
• May 20 : Jan open-sourced
• May 17 : OpenAlpha_Evolve
• May 21 : Devstral
• May 23 : QwenLong-L1
• May 28 : Deepseek R1-0528
• May 29 : Skywork-OR1
• May 30 : MiMo

June

• Jun. 2 : AlphaOne
• Jun. 5 : The Common Pile v0.1
• Jun. 6 : Dots.llm1
• Jun. 6 : OpenThoughts3
• Jun. 6 : Seed-Coder
• Jun. 9 : Nano-vLLM
• Jun. 10 : Magistral-Small-2506
• Jun. 12 : NextCoder
• Jun. 16 : MiniMax-M1
• Jun. 16 : OmniGen2
• Jun. 20 : Mistral-Small-3.2
• Jun. 23 : Google donates A2A to Linux Foundation
• Jun. 25 : Google Gemini CLI
• Jun. 26 : Google Gemma 3n
• Jun. 26 : Kwai Keye-VL
• Jun. 27 : Hunyuan-A13B
• Jun. 29 : ERNIE from Baidu

July

• Jul. 1 : DeepSWE-Preview
• Jul. 1 : DiffuCoder-7B-cpGRPO
• Jul. 2 : DeepSeek-TNG-R1T2-Chimera
• Jul. 6 : MemOS
• Jul. 8 : SmolLM3
• Jul. 9 : Richy Mini by Pollen Robotics
• Jul. 9 : Alps LLM
• Jul. 11 : Kimi K2
• Jul. 14 : Abandoning Meta's most powerful open source A.I. model in favor of developing a closed one
• Jul. 18 : Seed-X
• Jul. 21 : Qwen3-235B-A22B-Instruct-2507
• Jul. 21 : Cogito v2
• Jul. 23 : US AI Action Plan encourages Open-Source and Open-Weight
• Jul. 25 : Chinese AI Models Register a 60 Percent Fail Rate in NewsGuard Audit
• Jul. 25 : China's Overlooked AI Strategy
• Jul. 26 : China : Global AI Governance Action Plan
• Jul. 28 : GLM-4.5

August

• Aug. 1 : XBai-o4
• Aug. 4 : ATOM : The American DeepSeek Project
• Aug. 5 : OpenAI Open Models
• Aug. 9 : The LibreModel Project (AMA)
• Aug. 12 : LFM2‐VL
• Aug. 14 : Gemma 3
• Aug. 17 : Qwen-Image-Edit
• Aug. 18 : Nemotron Nano
• Aug. 19 : DeepSeek-V3.1
• Aug. 20 : ByteDance's Seed-OSS
• Aug. 22 : xAI's Grok-2
• Aug. 24 : The best non-Chinese model is GPT OSS 120B, ranked at 16th
• Aug. 25 : Hermes 4
• Aug. 27 : Hugging Face has reached two million models

September

• Sep. 2 : Swiss Apertus
• Sep. 4 : EmbeddingGemma
• Sep. 8 : UAE's K2-Think
• Sep. 8 : ERNIE-4.5
• Sep. 17 : Magistral Small 1.2
• Sep. 18 : Tongyi DeepResearch
• Sep. 19 : DeepSeek cost just 294,000ドル to train
• Sep. 11 : The open-source AI ecosystem is navigating a legal minefield
• Sep. 22 : Qwen3-Omni
• Sep. 22 : The great schism — Where does AI go from here?
• Sep. 24 : Meituan LongCat
• Sep. 25 : Liquid Nanos
• Sep. 26 : Models released or udpated last week
• Sep. 29 : A Cartography of Open Collaboration in Open Source AI

October

• Oct. 1 : IBM Granite 4.0
• Oct. 8 : EU Apply AI Strategy
• Oct. 11 : Ant group Ring-1T
• Oct. 18 : Open Source AI Week
• Oct. 20 : DeepSeek-OCR
• Oct. 23 : MiniMax-M2
• Oct. 23 : Granite 4.0 Nano
• Oct. 29 : gpt-oss-safeguard
• Oct. 29 : OpenAI lost more than 11ドル.5B last quarter
• Oct. 31 : Cursor new models are speculated to be built on Chinese base models
• Oct. 31 : Open-source models released this month

November

• Nov. 3 : The State of AI: is China about to win the race?
• Nov. 6 : Meta relied on scam ad profits to fund AI
• Nov. 6 : Nvidia's Jensen Huang: 'China is going to win the AI race,'
• Nov. 6 : Alibaba-backed Moonshot releases its second AI update in four months as China’s AI race heats up
• Nov. 6 : Is this another DeepSeek moment?
• Nov. 6 : 5 Thoughts on Kimi K2 Thinking
• Nov. 7 : World's strongest agentic model is now open source
• Nov. 7 : Kimi K2 Thinking Crushes GPT-5, Claude 4.5 Sonnet in Key Benchmarks
• Nov. 10 : ERNIE-4.5-VL-28B-A3B-Thinking
• Nov. 10 : VibeThinker-1.5B (7,800ドル USD training cost)
• Nov. 14 : Why the Future of AI Must Be Built in the Open
• Nov. 19 : Cogito v2
• Nov. 20 : Olmo 3
• Nov. 23 : A resurgence of truly open models
• Nov. 24 : White House's Genesis Mission
• Nov. 26 : China leapfrogs US in global market for ‘open’ AI models
• Nov. 27 : China’s tech giants take AI model training offshore
• Nov. 27 : Has the bailout of generative AI already begun?
• Nov 27 : Economies of Open Intelligence

December

• Dec. 1 : Arcee Trinity
• Dec. 1 : DeepSeek-V3.2
• Dec. 2 : Mistral 3
• Dec. 4 : The US may be running the wrong AI race
• Dec. 5 : Essential AI
• Dec. 7 : Why the Model Context Protocol Won
• Dec. 8 : Open-weight models are generally three to six months behind
• Dec. 9 : China’s open-source AI is a national advantage
• Dec. 9 : Devstral2
• Dec. 9 : Agentic AI Foundation (AAIF)
• Dec. 9 : Donating the Model Context Protocol and establishing AAIF
• Dec. 10 : DeepSeek is Using Banned Nvidia Chips
• Dec. 11 : Nvidia verification tech against chip smuggling
• Dec. 14 : Molmo 2
• Dec. 14 : EuroLLM 22b
• Dec. 15 : Nemotron 3 Nano
• Dec. 15 : QwenLong-L1.5
• Dec. 15 : Why China Can’t Win the AI-Led Industrial Revolution

Contact

• Have questions ?
• Want to contribute ?
• Submit a Pull Request
• Or write to contact at fordj dot org