Person: "simonw"

Kimi K2 - SOTA Open MoE proves that Muon can scale to 15T tokens/1T params

kimi-k2 kimi-k2-1t deepseek-v3 grok-4 devstral-2507 gpt-4.1 sonnet-4 moonshot-ai alibaba tencent deepseek x-ai mistral-ai weights-biases hugging-face mixture-of-experts model-training model-optimization optimizer benchmarking long-context model-performance open-weights model-release yuchenj_uw andrew_n_carr scaling01 novita_labs teknium1 aravsrinivas mparakhin simonw

Moonshot AI has released Kimi K2, a 1 trillion parameter Mixture-of-Experts model trained on 15.5 trillion tokens using the new MuonClip optimizer, achieving state-of-the-art results on benchmarks like SWE-Bench Verified (65.8%) and TAU2 (58.4%). This model is competitive with GPT-4.1 and Sonnet 4 on non-thinking tasks and is available under an MIT license. Meanwhile, xAI announced Grok-4, noted for its "LEAST censored frontier model" status and strong long-context performance but criticized for rushed post-training. Mistral AI updated its Devstral 2507 models with improved performance and cost efficiency. The community is excited about the potential of the MuonClip optimizer, which may surpass the long-standing AdamW optimizer in machine learning.

Jun 27

not much happened today

gemma-3n hunyuan-a13b flux-1-kontext-dev mercury fineweb2 qwen-vlo o3-mini o4-mini google-deepmind tencent black-forest-labs inception-ai qwen kyutai-labs openai langchain langgraph hugging-face ollama unslothai nvidia amd multimodality mixture-of-experts context-windows tool-use coding image-generation diffusion-models dataset-release multilinguality speech-to-text api prompt-engineering agent-frameworks open-source model-release demishassabis reach_vb tri_dao osanseviero simonw clementdelangue swyx hwchase17 sydneyrunkle

Google released Gemma 3n, a multimodal model for edge devices available in 2B and 4B parameter versions, with support across major frameworks like Transformers and Llama.cpp. Tencent open-sourced Hunyuan-A13B, a Mixture-of-Experts (MoE) model with 80B total parameters and a 256K context window, optimized for tool calling and coding. Black Forest Labs released FLUX.1 Kontext [dev], an open image AI model gaining rapid Hugging Face adoption. Inception AI Labs launched Mercury, the first commercial-scale diffusion LLM for chat. The FineWeb2 multilingual pre-training dataset paper was released, analyzing data quality impacts. The Qwen team released Qwen-VLo, a unified visual understanding and generation model. Kyutai Labs released a top-ranked open-source speech-to-text model running on Macs and iPhones. OpenAI introduced Deep Research API with o3/o4-mini models and open-sourced prompt rewriter methodology, integrated into LangChain and LangGraph. The open-source Gemini CLI gained over 30,000 GitHub stars as an AI terminal agent.

May 27

Mistral's Agents API and the 2025 LLM OS

qwen claude-4 chatgpt o3 o4 mistral-ai langchain-ai openai meta-ai-fair agent-frameworks multi-agent-systems tool-use code-execution web-search model-context-protocol persistent-memory function-calling open-source no-code reinforcement-learning model-performance agent-orchestration omarsar0 simonw swyx scaling01

The LLM OS concept has evolved since 2023, with Mistral AI releasing a new Agents API that includes code execution, web search, persistent memory, and agent orchestration. LangChainAI introduced the Open Agent Platform (OAP), an open-source no-code platform for intelligent agents. OpenAI plans to develop ChatGPT into a super-assistant by H1 2025, competing with Meta. Discussions around Qwen models focus on reinforcement learning effects, while Claude 4 performance is also noted. The AI Engineer World's Fair is calling for volunteers.

Dec 18, 2024

o1 API, 4o/4o-mini in Realtime API + WebRTC, DPO Finetuning

o1-2024-12-17 o1 o1-pro 4o 4o-mini gemini-2-0-flash claude-3.5-sonnet claude-3.5 openai google google-deepmind function-calling structured-outputs vision reasoning webrtc realtime-api preference-tuning fine-tuning api model-performance aidan_mclau kevinweil simonw michpokrass morgymcg juberti

OpenAI launched the o1 API with enhanced features including vision inputs, function calling, structured outputs, and a new reasoning_effort parameter, achieving 60% fewer reasoning tokens on average. The o1 pro variant is confirmed as a distinct implementation coming soon. Improvements to the Realtime API with WebRTC integration offer easier usage, longer sessions (up to 30 minutes), and significantly reduced pricing (up to 10x cheaper with mini models). DPO Preference Tuning for fine-tuning is introduced, currently available for the 4o model. Additional updates include official Go and Java SDKs and OpenAI DevDay videos. The news also highlights discussions on Google Gemini 2.0 Flash model's performance reaching 83.6% accuracy.