All tags
Company: "vllm_project"
not much happened today
nemotron-3-ultra nemotron-3.5-asr claude-opus-4 mythos-preview nvidia anthropic togethercompute baseten modal vllm_project fireworksai_hq ollama wandb cline primeintellect nousresearch mixture-of-experts long-context model-quantization agentic-ai streaming-speech asr low-precision-training benchmarking recursive-self-improvement code-generation model-speedup piotrz_zelasko
NVIDIA released Nemotron 3 Ultra, a fully open 550B MoE model with 55B active parameters and 1M context, optimized for long-running agent tasks with up to 5x speedup and 30% cost reduction. It features hybrid Mamba/attention, LatentMoE, native MTP, and was pretrained on 20T tokens using NVFP4 low-precision format. Benchmarks show strong performance with 47.7 Intelligence Index and 400+ output tokens/sec. The model is supported across major serving platforms. Additionally, Nemotron 3.5 ASR is an open streaming ASR model with 0.6B parameters, supporting 40 language-locale combinations and sub-100ms latency, designed for voice agents.
Anthropic highlighted early signs of recursive self-improvement (RSI) in AI, with Claude models authoring 80%+ of merged code and engineers shipping 8x more code. Claude Opus 4 achieved 3x speedup on training scripts, while Mythos Preview reached ~52x speedup and provided better research suggestions than humans 64% of the time.
not much happened today
claude-opus-4.8 gpt-5.5 qwen kimi deepseek anthropic huggingface langchain vllm_project reinforcement-learning tokenization agentic-ai api model-optimization long-context rust performance-optimization multi-agent-systems prompt-engineering jeremyphoward leo_linsky clementdelangue johnschulman2 omarsar0 hwchase17 ofirpress scaling01
Anthropic rolled out Claude Opus 4.8, which shows incremental improvements but mixed benchmark results, including better cooperation and coding behavior but some regressions in document parsing. Platform updates include mid-conversation system instructions enhancing long agent sessions, though API pricing remains a concern. A Hugging Face analysis revealed a critical bug in multi-turn reinforcement learning training loops involving tokenization mismatches, with a proposed "Token-In, Token-Out" fix. Agent harness design is evolving as a key optimization area, with LangChain's Deep Agents v0.6 achieving strong performance at much lower cost, and vllm_project releasing native weight syncing APIs and a Rust BPE tokenizer to improve tokenization efficiency. Debate continues on the value of multi-agent systems, with some seeing them as speedups and others expecting capability breakthroughs.
not much happened today
eagle-3.1 unigram-tokenizer qwen-3.5 deepseek-v4-pro mimo deep-agents-v0.6 397b-parameter-model eaglecorp vllm_project perplexity_ai alibaba lightseek nvidia mooncake flashattention kimmonismus deepseek xiaomi langchain baseten trajectory clay harvey decagon mercor rogo rlm inference-optimization long-context speculative-decoding tokenization attention-mechanisms kv-cache cache-hierarchy agent-engineering model-harness-memory-fit continual-learning quantization autoscaling memory-centric-agents evaluation-automation kimmonismus _luofuli vtrivedy10
Inference optimization is increasingly architectural, with EAGLE 3.1 improving speculative decoding and long-context handling, collaborating with vLLM and TorchSpec. Perplexity open-sourced a rebuilt Unigram tokenizer cutting CPU use by 5–6× and achieving 63 µs at 514 tokens. Qwen3.5 hits 580 tokens/s via joint efforts from Alibaba, LightSeek, NVIDIA, Mooncake, and FlashAttention-4 contributors. Price cuts in APIs from Chinese labs are sustainable due to structural KV-cache and attention improvements, exemplified by DeepSeek V4-Pro and Xiaomi MiMo reducing caching costs significantly.
Agent engineering shifts focus from model quality to model-harness-memory fit, with LangChain releasing Deep Agents v0.6 and tools like LangSmith Engine automating evaluation loops. Trajectory launched a continual learning platform with $15M funding and partners like Clay and Harvey, supporting large models including a 397B-parameter model deployed on autoscaled H100 infrastructure. Open-source memory-centric agents and minimal training harnesses also gained attention.
not much happened today
gpt-5.5 gpt-image-2 gpt-5.5-pro gpt-5.5-instant gpt-realtime-2 gpt-5.5-cyber codex zaya1-74b-preview zaya1-vl-8b qwen3-omni openai zyphra amd deepseek vllm_project model-release model-training mixture-of-experts inference model-optimization sandboxing alignment cybersecurity agent-runtime throughput quantization telemetry real-time-detection reach_vb dhh gdb patience_cave ithilgore cryps1s sama deredleritt3r
OpenAI rapidly expanded the GPT-5.5 family with multiple variants including gpt-image-2, GPT-5.5 Pro, and GPT-5.5 Cyber, receiving positive feedback for efficiency and usability. Codex evolved into a long-running agent runtime with a new /goal mechanism, achieving 61% success on ARC-AGI-3 games after extensive testing. OpenAI also introduced cybersecurity-focused models like GPT-5.5-Cyber targeting enterprise and government sectors. Meanwhile, Zyphra released the open-model ZAYA1-74B-Preview, a 74B parameter mixture-of-experts model trained on AMD hardware under Apache 2.0 license, alongside a vision-language model ZAYA1-VL-8B. Inference infrastructure competition intensified with vLLM updates improving throughput and latency, including support for DeepSeek V4 and enhanced quantization/backends.
not much happened today
gpt-5.5 gpt-5.4 opus-4.7 mimo-v2.5-pro mimo-v2.5 kimi-k2.6 codex copilot openai microsoft google amazon github xiaomi openai-devs vllm_project kimi-moonshot model-distribution cloud-computing benchmarking usage-based-billing model-orchestration open-source large-context-models agent-scaling coding model-training fp8 attention-mechanisms multi-agent-systems sama scaling01 kimmonismus ajassy simonw htihle arena gdb hangsiin eliebakouch _luofuli teortaxestex
OpenAI loosens its Azure exclusivity, allowing distribution across Google TPU, AWS Trainium, and Bedrock with commitments through 2032 and revenue share through 2030. GPT-5.5 shows improved benchmarks but is not uniformly dominant, ranking variably across coding, document, math, and vision tasks. GitHub's Copilot shifts to usage-based billing starting June 1, reflecting increased runtime costs. OpenAI open-sourced Symphony, an orchestration layer for issue tracking and Codex agents. Xiaomi released MiMo-V2.5 and MiMo-V2.5-Pro, large context models with up to 1M-token context and trillions of tokens trained, emphasizing complex agent and omni-modal capabilities. Kimi K2.6 leads OpenRouter's leaderboard, noted for coding and long-horizon agent capabilities with large-scale sub-agent coordination.
not much happened today
qwen3.6-27b qwen3.5-397b-a17b privacy-filter mimo-v2.5-pro mimo-v2.5 gemini-3.1-pro gemini-3.1-flash-image alibaba openai xiaomi google google-deepmind vllm_project unsloth ggml ollama arena nous-research open-models multimodality vision tokenization pii-detection privacy enterprise-ai agentic-ai benchmarking long-context model-deployment hardware-optimization model-integration software-engineering alibaba_qwen clementdelangue altryne eliebakouch mervenoyann xiaomimo sundarpichai scaling01
Alibaba released Qwen3.6-27B, a dense, Apache 2.0 open coding model with thinking and non-thinking modes, outperforming the larger Qwen3.5-397B-A17B on multiple coding benchmarks including SWE-bench and Terminal-Bench. It supports native vision-language reasoning over images and video, with immediate ecosystem support from vLLM, Unsloth, ggml, and Ollama. OpenAI open-sourced a practical Privacy Filter model for PII detection and masking, a 1.5B parameter token-classification model with a 128k context window aimed at enterprise redaction tasks. Xiaomi announced MiMo-V2.5-Pro and MiMo-V2.5 models, emphasizing software engineering advances, long-horizon agents, and large context windows (up to 1M tokens), with strong benchmark results and integrations with Hermes and Nous. At Google Cloud Next, Google and Google DeepMind unveiled 8th-gen TPUs (TPU 8t for training and TPU 8i for inference) with claims of scaling to a million TPUs in a cluster, and launched the Gemini Enterprise Agent Platform evolving Vertex AI with Agent Studio and access to 200+ models including Gemini 3.1 Pro and Gemini 3.1 Flash Image. This marks a significant vertical integration of hardware, models, and enterprise tooling.
not much happened today
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."
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.
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.
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.
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 Fran ois 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.