All tags
Company: "baseten"
MiniMax M2 230BA10B — 8% of Claude Sonnet's price, ~2x faster, new SOTA open model
minimax-m2 hailuo-ai huggingface baseten vllm modelscope openrouter cline sparse-moe model-benchmarking model-architecture instruction-following tool-use api-pricing model-deployment performance-evaluation full-attention qk-norm gqa rope reach_vb artificialanlys akhaliq eliebakouch grad62304977 yifan_zhang_ zpysky1125
MiniMax M2, an open-weight sparse MoE model by Hailuo AI, launches with ≈200–230B parameters and 10B active parameters, offering strong performance near frontier closed models and ranking #5 overall on the Artificial Analysis Intelligence Index v3.0. It supports coding and agent tasks, is licensed under MIT, and is available via API at competitive pricing. The architecture uses full attention, QK-Norm, GQA, partial RoPE, and sigmoid routing, with day-0 support in vLLM and deployment on platforms like Hugging Face and Baseten. Despite verbosity and no tech report, it marks a significant win for open models.
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
gemini-robotics-1.5 gemini-live embeddinggemma veo-3 gemini-2.5-flash code-world-model-32b qwen3-coder-30b vllm-v1 mlx-lm flashattention-4 google meta-ai-fair perplexity-ai baseten spatial-reasoning temporal-reasoning agentic-ai code-semantics code-execution-traces coding-infrastructure runtime-optimization batch-inference embedding-latency api model-optimization model-performance osanseviero _anniexie rmstein scaling01 giffmana cline redhat_ai awnihannun charles_irl bernhardsson akshat_b aravsrinivas
Google released a dense September update including Gemini Robotics 1.5 with enhanced spatial/temporal reasoning, Gemini Live, EmbeddingGemma, and Veo 3 GA powering creative workflows. They also introduced agentic features like restaurant-reservation agents and reduced pricing for Gemini 2.5 Flash. Meta AI unveiled the open-weight Code World Model (CWM) 32B, excelling in code semantics and math benchmarks, with innovations in training code models via execution traces. Local-first coding setups highlight Qwen3-Coder-30B running efficiently on consumer GPUs, paired with tools like Cline and LM Studio. Runtime improvements include vLLM v1 supporting hybrid models and mlx-lm adding batch inference on Apple silicon. In infrastructure, FlashAttention 4 was reverse-engineered revealing a ~20% speedup from architectural optimizations. Perplexity AI advances its independent web index and browsing API with upcoming feed refreshes. Embedding latency improvements were achieved by Superhuman using Baseten.
Qwen3-Next-80B-A3B-Base: Towards Ultimate Training & Inference Efficiency
qwen3-next qwen3 mixtral-8x7b gemini-2.5-pro alibaba mistral-ai deepseek snowflake hugging-face baseten nvidia mixture-of-experts model-sparsity gated-attention hybrid-architecture rmsnorm model-stability model-training inference-optimization multi-token-prediction model-deployment justinlin610 teortaxestex yuchenj_uw
MoE (Mixture of Experts) models have become essential in frontier AI models, with Qwen3-Next pushing sparsity further by activating only 3.7% of parameters (3B out of 80B) using a hybrid architecture combining Gated DeltaNet and Gated Attention. This new design includes 512 total experts (10 routed + 1 shared), Zero-Centered RMSNorm for stability, and improved MoE router initialization, resulting in ~10× cheaper training and 10× faster inference compared to previous models. Alibaba's Qwen3-Next reportedly outperforms Gemini-2.5-Flash-Thinking and approaches the flagship 235B model's performance, with deployments on Hugging Face, Baseten, and native vLLM support for efficient inference.
not much happened today
qwen-image-edit qwen-vl-max kling-2.1 veo-3 deepseek-v3.1 genie-3 sima google-deepmind alibaba google deepseek baseten yupp multimodality embodied-ai simulation fine-tuning quantization video-generation image-generation local-inference scaling agent-training real-time-control spatial-memory demishassabis bonniesjli shreyar ostrisai lmarena_ai teortaxestex ivanfioravanti
DeepMind released Genie 3, an interactive multimodal world simulator with advanced spatial memory and real-time avatar control, and SIMA, an embodied training agent operating inside generated worlds. Alibaba introduced Qwen-Image-Edit, an open-weights image editor scoring ELO 1098 (#2) in the Image Editing Arena, running on Qualcomm NPUs, alongside Qwen-VL-Max entering the Vision top-20. Video models like Kling 2.1 showed a 235% improvement in frame control, with new entrants Luma Ray 2 and Runway Gen-4 Turbo debuting. Google provided free Veo 3 generations in Gemini App and enhanced Google Photos with natural-language edits. DeepSeek v3.1 launched with focus on SWE and Search agents, supporting local inference on Apple Silicon with 4-bit quantization achieving ~21 tok/s on M3 Ultra. The news highlights advances in interactive simulation, vision editing, video synthesis, and scalable local AI inference.
Cohere Command A Reasoning beats GPT-OSS-120B and DeepSeek R1 0528
command-a-reasoning deepseek-v3.1 cohere deepseek intel huggingface baseten vllm-project chutes-ai anycoder agentic-ai hybrid-models long-context fp8-training mixture-of-experts benchmarking quantization reasoning coding-workflows model-pricing artificialanlys reach_vb scaling01 cline ben_burtenshaw haihaoshen jon_durbin _akhaliq willccbb teortaxestex
Cohere's Command A Reasoning model outperforms GPT-OSS in open deep research capabilities, emphasizing agentic use cases for 2025. DeepSeek-V3.1 introduces a hybrid reasoning architecture toggling between reasoning and non-reasoning modes, optimized for agentic workflows and coding, with extensive long-context pretraining (~630B tokens for 32k context, ~209B for 128k), FP8 training, and a large MoE expert count (~37B). Benchmarks show competitive performance with notable improvements in SWE-Bench and other reasoning tasks. The model supports a $0.56/M input and $1.68/M output pricing on the DeepSeek API and enjoys rapid ecosystem integration including HF weights, INT4 quantization by Intel, and vLLM reasoning toggles. Community feedback highlights the hybrid design's pragmatic approach to agent and software engineering workflows, though some note the lack of tool use in reasoning mode.
DeepSeek V3.1: 840B token continued pretrain, beating Claude 4 Sonnet at 11% of its cost
deepseek-v3.1 seed-oss-36b computerrl gemini-2.5-pro gpt-5 claude-code gpt-oss-120b gpt-oss-20b deepseek bytedance zhipu-ai github microsoft anthropic together-ai baseten huggingface token-efficiency coding agentic-benchmarks long-context reinforcement-learning developer-tools fine-tuning multinode-training model-release teortaxestex rasbt lukehoban burkeholland _catwu cline winglian
DeepSeek released DeepSeek V3.1, a quietly rolled out open model with an 128K context window and improvements in token efficiency, coding, and agentic benchmarks. ByteDance launched the permissive Seed-OSS 36B model on Hugging Face, noted for long-context and reasoning capabilities. Zhipu AI introduced ComputerRL, a reinforcement learning framework for computer-use agents, achieving strong benchmark results. In developer tooling, GitHub Copilot expanded globally, Microsoft VS Code integrated Gemini 2.5 Pro and updated GPT-5 agent prompts, and Anthropic launched Claude Code seats with spend controls. Open-source fine-tuning advances include Together AI adding SFT for gpt-oss-120B/20B and Baseten enabling multinode 120B training with Truss CLI. The community noted mixed performance and ongoing post-training adjustments for DeepSeek V3.1.
not much happened today
gpt-oss-120b gpt-oss-20b kimi-k2 deepseek-r1 qwen-3-32b openai huggingface microsoft llamaindex ollama baseten fireworksai cerebras groq together anthropic google uk-aisi sliding-window-attention mixture-of-experts rope context-length mxfp4-format synthetic-data reasoning-core-hypothesis red-teaming benchmarking coding-benchmarks model-performance fine-tuning woj_zaremba sama huybery drjimfan jxmnop scaling01 arunv30 kevinweil xikun_zhang_ jerryjliu0 ollama basetenco reach_vb gneubig shxf0072 _lewtun
OpenAI released its first open models since GPT-2, gpt-oss-120b and gpt-oss-20b, which quickly trended on Hugging Face. Microsoft supports these models via Azure AI Foundry and Windows Foundry Local. Key architectural innovations include sliding window attention, mixture of experts (MoE), a RoPE variant, and a 256k context length. The models use a new MXFP4 format supported by llama.cpp. Hypotheses suggest gpt-oss was trained on synthetic data to enhance safety and performance, supporting the Reasoning Core Hypothesis. OpenAI announced a $500K bounty for red teaming with partners including Anthropic, Google, and the UK AISI. Performance critiques highlight inconsistent benchmarking results, with GPT-OSS-120B scoring 41.8% on the Aider Polyglot coding benchmark, trailing competitors like Kimi-K2 and DeepSeek-R1. Some users note the model excels in math and reasoning but lacks common sense and practical utility.
Ideogram 2 + Berkeley Function Calling Leaderboard V2
llama-3-70b gpt-4 phi-3.5 functionary-llama-3-70b llama-3 ideogram midjourney berkeley openai hugging-face microsoft meta-ai-fair baseten kai claude functionary function-calling benchmarking image-generation model-optimization vision multimodality model-performance fine-tuning context-windows cybersecurity code-analysis ai-assisted-development
Ideogram returns with a new image generation model featuring color palette control, a fully controllable API, and an iOS app, reaching a milestone of 1 billion images created. Meanwhile, Midjourney released a Web UI but still lacks an API. In function calling, the Berkeley Function Calling Leaderboard (BFCL) updated to BFCL V2 • Live, adding 2251 live, user-contributed function documentation and queries to improve evaluation quality. GPT-4 leads the leaderboard, but the open-source Functionary Llama 3-70B finetune from Kai surpasses Claude. On AI model releases, Microsoft launched three Phi-3.5 models with impressive reasoning and context window capabilities, while Meta AI FAIR introduced UniBench, a unified benchmark suite for over 50 vision-language model tasks. Baseten improved Llama 3 inference speed by up to 122% using Medusa. A new cybersecurity benchmark, Cyberbench, featuring 40 CTF tasks, was released. Additionally, Codegen was introduced as a tool for programmatic codebase analysis and AI-assisted development. "Multiple functions > parallel functions" was highlighted as a key insight in function calling.
ALL of AI Engineering in One Place
claude-3-sonnet claude-3 openai google-deepmind anthropic mistral-ai cohere hugging-face adept midjourney character-ai microsoft amazon nvidia salesforce mastercard palo-alto-networks axa novartis discord twilio tinder khan-academy sourcegraph mongodb neo4j hasura modular cognition anysphere perplexity-ai groq mozilla nous-research galileo unsloth langchain llamaindex instructor weights-biases lambda-labs neptune datastax crusoe covalent qdrant baseten e2b octo-ai gradient-ai lancedb log10 deepgram outlines crew-ai factory-ai interpretability feature-steering safety multilinguality multimodality rag evals-ops open-models code-generation gpus agents ai-leadership
The upcoming AI Engineer World's Fair in San Francisco from June 25-27 will feature a significantly expanded format with booths, talks, and workshops from top model labs like OpenAI, DeepMind, Anthropic, Mistral, Cohere, HuggingFace, and Character.ai. It includes participation from Microsoft Azure, Amazon AWS, Google Vertex, and major companies such as Nvidia, Salesforce, Mastercard, Palo Alto Networks, and more. The event covers 9 tracks including RAG, multimodality, evals/ops, open models, code generation, GPUs, agents, AI in Fortune 500, and a new AI leadership track. Additionally, Anthropic shared interpretability research on Claude 3 Sonnet, revealing millions of interpretable features that can be steered to modify model behavior, including safety-relevant features related to bias and unsafe content, though more research is needed for practical applications. The event offers a discount code for AI News readers.