All tags
Company: "modal"
not much happened today
glm-5.3 qwen3.8-27b qwen3.8-2.4t-a95b deepseek-v4-pro dots3-note z-ai alibaba deepseek rednote vllm together-ai fireworks modal digitalocean deepinfra unsloth post-training reinforcement-learning agent-runtimes long-horizon-training multimodality model-infrastructure runtime-architecture model-benchmarking open-weight apache-2.0-license model-optimization multimodal-models mixture-of-experts context-windows
Z.ai launched GLM-5.3, a coding- and cyber-focused model with significant gains on agentic and security benchmarks, achieved through scaled post-training rather than a larger base model. Alibaba released Qwen3.8-27B, a native multimodal dense model under Apache 2.0 with a 262K native context extendable to 1M, designed for real-world coding and office workflows, with broad inference support from multiple platforms. DeepSeek V4-Pro and RedNote's dots3-note, a 280B multimodal MoE model with 16B active parameters and 512K context, continue the China open-model wave, introducing new RL methods like TEMPO for long-horizon self-evaluation. The ecosystem features multiple Chinese labs specializing in open models with different strengths. DeepSeek's harness is highlighted as a modular agent runtime infrastructure with replaceable components and lifecycle management via Cordis.
not much happened today
gpt-5.6-luna gpt-5.6-terra gpt-5.6-sol arc-agi-3 inkling-small inkling gemini-robotics-2 openai thinking-machines lmsys modal unsloth artificial-analysis google price-optimization agent-systems memory-retention context-compaction multimodality mixture-of-experts model-compression benchmarking open-weights multimodal-models model-efficiency model-deployment embodied-ai robotics long-context sama fchollet kimmonismus gneubig scaling01 mervenoyann
OpenAI aggressively cut prices for GPT-5.6 Luna by 80% and Terra by 20%, introducing a faster Sol Fast tier with up to 2.5× lower latency at double the price, improving agent workflow costs by roughly 10×. The ARC-AGI-3 debate highlighted that the complete agent system, including memory retention and tool orchestration, is critical beyond just the base model. Thinking Machines released Inkling-Small, an open-weights, multimodal MoE model with 276B parameters (12B active), delivering performance comparable to the original Inkling at a quarter of the size, supporting audio, images, and Python-based image inspection. Benchmarks show Inkling-Small excels in coding and multimodality tasks, with 1M-context support and broad open inference stack adoption. The news also mentions Google's Gemini Robotics 2 advancing embodied AI from tabletop to full-body control.
not much happened today
kimi-k3 moonshot vllm baseten modal together-ai ollama dell nvidia mixture-of-experts model-scaling numerical-stability model-architecture open-models model-distribution model-licensing agentic-ai vision scaling-efficiency open-source-infrastructure commercial-restrictions ai-security kimi_moonshot jensenhuang natolambert petergostev artificialanlys
Moonshot released the Kimi K3 open-weights model, a 2.8T-parameter MoE with 104B active parameters, 896 experts, and 1M-token context featuring native visual understanding. The release includes open-source infrastructure like FlashKDA, MoonEP, and AgentENV, enabling large-scale agentic post-training and serving. The technical report highlights a ~2.5× scaling-efficiency improvement over K2 with innovations in numerical stability and MoE routing. Licensing is source-available with commercial-use restrictions, signaling a trend towards open-weight models with business carve-outs. Distribution was broad and immediate via platforms like vLLM, Baseten, Modal, Together, and Ollama Cloud. Separately, NVIDIA launched the Open Secure AI Alliance to build an ecosystem combining open and closed frontier models for AI security, emphasizing defense against attackers already equipped with strong AI.
not much happened today
inkling thinking-machines-lab huggingface vllm_project lmsysorg modal baseten databricks mixture-of-experts multimodality foundation-models model-licensing context-window open-weights model-release miramurati soumithchintala johnschulman2 lilianweng natolambert artificialanlys scaling01
Thinking Machines Lab launched Inkling, its first fully released open-weights foundation model family, featuring 975B parameters with 41B active parameters in a Mixture-of-Experts architecture. Inkling supports multimodality with text, image, and audio inputs and text output, is Apache 2.0 licensed, and offers up to 1M context window. The model is available on platforms like Tinker, Hugging Face, and partners, with broad ecosystem support from vLLM, SGLang, Modal, Baseten, and Databricks. Key figures such as Mira Murati, Soumith Chintala, John Schulman, and Lilian Weng highlighted its open weights, customization, and practical use focus. Independent commentators noted it as the strongest U.S.-based open-weight release to date, though still behind top Chinese open-weight and best closed models on some benchmarks.
not much happened today
dflash nemo-automodel claude openai broadcom qualcomm modular nvidia skypilot modal anthropic hugging-face hardware inference performance-optimization model-training agent-ux security capability-based-security open-source fine-tuning infrastructure model-optimization gdb kimmonismus scaling01 clattner_llvm karpathy gallabytes dabit3 kentonvarda random_walker jubbaonjeans victormustar
OpenAI announced Jalapeño, its first custom AI chip for LLM inference, built with Broadcom, aiming to control more of the AI stack and improve compute economics with a fast 9-month design cycle. Community analysis suggests Jalapeño features 216GB HBM3E, ~7.1–7.4 TB/s bandwidth, and ~10 PFLOPS FP4 performance, signaling hyperscaler-style inference silicon as a new standard. Meanwhile, Qualcomm is acquiring Modular, with Mojo open-sourcing on track, indicating rising competition in vertically integrated inference stacks beyond NVIDIA/CUDA. On infrastructure, NVIDIA's NeMo AutoModel boosts training throughput for MoE models by 3.4–3.7x, and startups like SkyPilot and Modal advance unified and open-source inference solutions. Custom training of DFLASH models yields 30–50% decode gains. In UX, Anthropic's Slack-native Claude agent shifts agent interaction from tools to coworkers, raising new security and cost concerns around identity, permissions, and lock-in, with debates on capability-based security and attribution. Hugging Face responded with its self-hosted Slack coding agent Moon Bot.
not much happened today
opus-4.8 gemma-4 cognition frontiercode moonshot google claudedevs magicpath langsmith modal coding-evaluation agent-control verification agent-ergonomics sandbox-environments local-inference workflow-optimization cli-tools plugin-integration persistent-memory swyx dzhng claudecode bcherny reach_vb omarsar0 gneubig hamelhusain angaisb_
FrontierCode benchmark by Cognition highlights the challenge of coding tasks with the best model, Opus 4.8, scoring only about 13% on the hardest subset, indicating coding is less solved than benchmarks suggest. The trend toward using loops as a control metaphor for coding agents is prominent, with emphasis on clear goals, verification, and iteration, though some experts caution about overreliance on loops. Agent ergonomics are improving with observability dashboards, sandbox environments, and workflow tools from ClaudeDevs, MagicPath, LangSmith, and Modal. Kimi by Moonshot released major updates including a stronger coding agent and a desktop agent product supporting up to 300 local sub-agents. Google advanced efficient local deployment with upgrades to Gemma 4 checkpoints.
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
OpenAI expanded its Agents SDK by separating the agent harness from compute/storage, enabling long-running, durable agents with features like file/computer use, skills, memory, and compaction. The harness is now open-source and supports execution via partner sandboxes, fostering a new ecosystem with integrations from Cloudflare, Modal, Vercel, and others. Cloudflare launched Project Think, a next-gen Agents SDK with durable execution and sandboxed code, alongside Agent Lee, a prompt-driven UI agent using sandboxed TypeScript, and introduced real-time voice pipelines and browser automation tools. Hermes Agent focuses on persistent skill formation by learning from completed workflows, positioning itself as a professional agent distinct from GUI-first assistants like OpenClaw. "Hermes autonomously backfills tracking data, updates cron jobs, and saves workflows as reusable skills," highlighting its advanced workflow management capabilities.
Z.ai GLM-5: New SOTA Open Weights LLM
glm-5 glm-4.5 kimi-k2.5 zhipu-ai openrouter modal deepinfra ollama qoder vercel deepseek-sparse-attention long-context model-scaling pretraining benchmarking office-productivity context-window model-deployment cost-efficiency
Zhipu AI launched GLM-5, an Opus-class model scaling from 355B to 744B parameters with DeepSeek Sparse Attention integration for cost-efficient long-context serving. GLM-5 achieves SOTA on BrowseComp and leads on Vending Bench 2, focusing on office productivity tasks and surpassing Kimi K2.5 on the GDPVal-AA benchmark. Despite broad availability on platforms like OpenRouter, Modal, DeepInfra, and Ollama Cloud, GLM-5 faces compute constraints impacting rollout and pricing. The model supports up to 200K context length and 128K max output tokens.
not much happened today.
gpt-5.2-codex glm-4.7 openai cursor github cerebras modal artificial-analysis vllm long-running-tasks autonomous-agents code-generation inference-speed latency batch-inference gpu-scaling model-evaluation agent-systems operational-scaling swyx kevinweil pierceboggan mntruell scaling01
OpenAI launched GPT-5.2-Codex API, touted as their strongest coding model for long-running tasks and cybersecurity. Cursor integrated GPT-5.2-Codex to autonomously run a browser for a week, producing over 3 million lines of Rust code. GitHub incorporated it into their code tools, easing enterprise adoption. Discussions highlight the importance of review loops in agent systems and debate evaluation metrics for coding models. OpenAI partnered with Cerebras to improve inference speed and latency, with Cerebras serving GLM-4.7 at 1,445 tokens/sec and low latency. Provider benchmarking reveals tradeoffs in throughput, latency, and context window sizes. Modal shared operational scaling insights for self-hosted inference fleets of 20k GPUs, focusing on batch inference optimization with vLLM and FlashInfer backend. This reflects a focus on inference infrastructure, long-horizon autonomous agents, and coding model evaluation.