Company: "vllm_project"

nomos-1 axiomprover devstral-2-small deepseek-v3.2 claude-code cursor-2.2 claude-opus-4.5 gpt-5 claude-sonnet-4.5 gemini-3-pro llama qwen mistral gemma nousresearch thinkymachines mistral-ai deepseek anthropic cursor microsoft langchain-ai openai gemini intel vllm_project danielhanchen math formal-reasoning agentic-systems asynchronous-execution multi-agent-systems observability benchmarking quantization post-training-quantization training-speedup kernel-optimization inference-efficiency

NousResearch's Nomos 1 is a 30B open math model achieving a top Putnam score with only ~3B active parameters, enabling consumer Mac inference. AxiomProver also posts top Putnam results using ThinkyMachines' RL stack. Mistral's Devstral 2 Small outperforms DeepSeek v3.2 in 71% of preferences with better speed and cost. Anthropic's Claude Code introduces asynchronous agent execution. Cursor 2.2 adds deep agent primitives like Debug and Plan Modes. VS Code launches unified agent chat sessions improving multi-agent workflows. LangChain releases "Polly" for agent observability. The Stirrup harness leads OpenAI GDPval benchmarks with Claude Opus 4.5, GPT-5, and Gemini 3 Pro following. Advances in quantization include vLLM integrating Intel's AutoRound PTQ for efficient serving. Unsloth achieves up to 3× training speedups with new kernels across Llama, Qwen, Mistral, and Gemma models. "Compositional reasoning + specialized post-training under constrained active params can rival frontier closed models on formal math."

Dec 08

not much happened today

glm-4.6v glm-4.6v-flash jina-vlm-2b hugging-face zhipu-ai jina-ai google-deepmind axiomprover fine-tuning multimodality model-optimization long-context mechanistic-interpretability formal-methods sequence-architectures reinforcement-learning lioronai akshay_pachaar _akhaliq ben_burtenshaw vllm_project prince_canuma zenmuxai eliebakouch theturingpost axiommathai neelnanda5 sarahookr

Claude Code Skills gains attention with a published talk and Hugging Face's new "skill" enabling one-line fine-tuning pipelines for models from ~0.5B to 70B parameters, supporting SFT, DPO, and GRPO, costing as low as ~$0.30 for small runs. Zhipu AI launches multimodal models GLM-4.6V (106B params MoE) and GLM-4.6V-Flash (9B dense), featuring 128k context and native multimodal function calling, with free Flash variant and API pricing detailed. Jina AI releases Jina-VLM (2B), a compact multilingual VLM excelling in diagrams and documents with top benchmark scores. At NeurIPS 2025, research highlights include Google's post-Transformer sequence architectures (Moneta, Yaad, Memora) showing up to 20% gains in long-context retrieval, AxiomProver's autonomous Lean system solving 9/12 Putnam 2025 problems rapidly, and mechanistic interpretability advances discussed by Chris Olah emphasizing scalable tooling.

Nov 06

Kimi K2 Thinking: 1T-A32B params, SOTA HLE, BrowseComp, TauBench && Soumith leaves Pytorch

kimi-k2-thinking gemini moonshot-ai google apple vllm_project arena baseten yupp_ai mixture-of-experts quantization int4 context-window agentic-ai benchmarking model-deployment inference-acceleration api performance-optimization eliebakouch nrehiew_ andrew_n_carr ofirpress artificialanlys sundarpichai akhaliq

Moonshot AI launched Kimi K2 Thinking, a 1 trillion parameter mixture-of-experts (MoE) model with 32 billion active experts, a 256K context window, and native INT4 quantization-aware training. It achieves state-of-the-art results on benchmarks like HLE (44.9%), BrowseComp (60.2%), and agentic tool use with 200-300 sequential tool calls. The model is deployed with vLLM support and OpenAI-compatible APIs, available on platforms like Arena, Baseten, and Yupp. Early user reports note some API instability under launch load. Meanwhile, Google announced the TPU v7 (Ironwood) with a 10× peak performance improvement over TPU v5p, aimed at training and agentic inference for models like Gemini. Apple added support for M5 Neural Accelerators in llama.cpp for inference acceleration.

Oct 24

not much happened today

nemotron-nano-2 gpt-oss-120b qwen3 llama-3 minimax-m2 glm-4.6-air gemini-2.5-flash gpt-5.1-mini tahoe-x1 vllm_project nvidia mistral-ai baseten huggingface thinking-machines deeplearningai pytorch arena yupp-ai zhipu-ai scaling01 stanford transformer-architecture model-optimization inference distributed-training multi-gpu-support performance-optimization agents observability model-evaluation reinforcement-learning model-provenance statistical-testing foundation-models cancer-biology model-fine-tuning swyx dvilasuero _lewtun clementdelangue zephyr_z9 skylermiao7 teortaxestex nalidoust

vLLM announced support for NVIDIA Nemotron Nano 2, featuring a hybrid Transformer–Mamba design and tunable "thinking budget" enabling up to 6× faster token generation. Mistral AI Studio launched a production platform for agents with deep observability. Baseten reported high throughput (650 TPS) for GPT-OSS 120B on NVIDIA hardware. Hugging Face InspectAI added inference provider integration for cross-provider evaluation. Thinking Machines Tinker abstracts distributed fine-tuning for open-weight LLMs like Qwen3 and Llama 3. In China, MiniMax M2 shows competitive performance with top models and is optimized for agents and coding, while Zhipu GLM-4.6-Air focuses on reliability and scaling for coding tasks. Rumors suggest Gemini 2.5 Flash may be a >500B parameter MoE model, and a possible GPT-5.1 mini reference appeared. Outside LLMs, Tahoe-x1 (3B) foundation model achieved SOTA in cancer cell biology benchmarks. Research from Stanford introduces a method to detect model provenance via training-order "palimpsest" with strong statistical guarantees.

Jul 23

not much happened today

qwen3-coder-480b-a35b-instruct kimi-k2 alibaba openrouterai togethercompute vllm_project unslothai white-house code-generation benchmarking model-integration context-windows open-source national-security infrastructure ai-policy fchollet clementdelangue scaling01 aravsrinivas rasbt gregkamradt yuchenj_uw

Alibaba announced the release of Qwen3-Coder-480B-A35B-Instruct, an open agentic code model with 480B parameters and 256K context length, praised for rapid development and strong coding performance. Benchmark claims of 41.8% on ARC-AGI-1 faced skepticism from Franois Chollet and others due to reproducibility issues. The model quickly integrated into ecosystems like vLLM, Dynamic GGUFs, and OpenRouterAI. The White House unveiled a new AI Action Plan emphasizing Innovation, Infrastructure, and International Diplomacy, linking AI leadership to national security and prioritizing compute access for the Department of Defense. The plan sparked debate on open vs. closed-source AI, with calls from Clement Delangue to embrace open science to maintain US AI competitiveness.