$ articles --topic=open-models
Open Models
63 AI articles about Open Models.
SageMaker Inference now supports Multi-Instance GPU to maximize GPU utilization for model endpointsExplains how Amazon SageMaker AI now uses NVIDIA Multi-Instance GPU to split one physical GPU into isolated partitions, letting you host more models per endpoint and choose right-sized GPU capacity. It covers the API fields AcceleratorPartitionLayouts and AcceleratorPartitionType, plus the trade-offs and measured results.Open Models · via AWS Machine Learning · 1d ago8 min[AINews] Inkling 2 Small: 17A-594B trained in 68 hours on 12T tokens, beating Muse SparkThis roundup covers Thinking Machines' Inkling 2 Small, a 594B-total, 17B-active multimodal MoE, including its benchmark gains over the predecessor, Tinker pricing, and the promised Apache 2.0 weights release. It also explains how the model compares with peers on coding, arenas, and domain evals, and the alignment problems early checkpoints showed.Evals · via Latent Space · 1d ago7 minZero Mismatch, Simple Async RL: REINFORCE Zero, DeepSeek-V4.1 and Kimi K3.1Explains how asynchronous reinforcement learning can train on stale model weights, and how aligning rollout and training kernels bit for bit, together with unclipped importance sampling, corrects that lag. Covers zero-mismatch RL applied to DeepSeek-V4.1-Flash, Kimi K3.1 and AMD MI355X hardware.Open Models · via vLLM · 1d ago1 minMigrating vLLM to the LibTorch Stable ABIExplains how the vLLM project moved its CUDA build and compiled extensions to PyTorch's Stable ABI. The write-up covers what changed in the build and why this lets vLLM work with newer PyTorch releases without a full rebuild.Open Models · via vLLM · 1d ago1 minFive Ways NVIDIA Co-Design Maximizes the Value of Every Byte in the AI FactoryExplains how NVIDIA's co-design approach gets more useful work out of high bandwidth memory in AI factories, covering the costs of redundant weight copies, idle KV caches, and recomputed prefixes. Describes five specific strategies and the generational performance data behind them.Open Models · via NVIDIA Technical Blog · 2d ago1 minLe Chonk brings the brains. Studio puts it to work.Learn what Mistral Studio offers for building and deploying enterprise agents, workflows, and apps alongside the Le Chonk model. The piece covers how teams can keep control over their models, data, and infrastructure while moving to production.Agents · via Mistral · 2d ago1 min[AINews] It's Swapping SeasonCovers the week's AI talent and company moves, including Cat Wu joining OpenAI, Sharon Zhou leaving AMD for Anthropic, and Meta acquiring Conductor. It also reviews OpenAI's autonomous pipeline that patched 948 Linux kernel privilege-escalation bugs and new open-weight models like Kimi K3.1 Pro and Ripple-65B.Open Models · via Latent Space · 2d ago8 minA Closer Look at Phonon TTS. Launching High-Quality On-Device Speech That Tops the Benchmarks.Open Models · via Hugging Face · 2d ago1 min[AINews] Behind Latent Space: Kill My SaaS, Jev & AI in 2026 | AIEi ShanghaiThis issue of AINews covers claims that frontier LLMs can read and write neural network weights directly, including a Qwen edit that lifted MATH-500 scores and its skeptical critiques. It also surveys new model releases, independent evals, provable inference from Anthropic, and audits that found flaws in popular benchmarks.Evals · via Latent Space · 3d ago7 minHow big is the open model economy today?The post asks how large the open model inference market is and whether it is growing or negligible next to OpenAI and Anthropic. It uses token volume disclosures from Baseten, Fireworks, Together AI, and OpenRouter's open model share, compared against Google's reported platform and API token figures, to ground the estimate.Open Models · via Interconnects · 3d ago3 mindatasette-llm 0.1a10Release notes for datasette-llm 0.1a10, a Datasette plugin for working with LLM models. It fixes a bug on the agent.datasette.io demo where the available models were not limited to those set in the configuration.Agents · via Simon Willison · 3d ago1 minvLLM Delivers Day-0 Support for Kimi K3.1 ProThis article explains how vLLM added day-0 support for Kimi K3.1 Pro. It covers the model's sparse attention on a KDA hybrid, its packed KV cache using FP8 or NVFP4 latents and MXFP4 index keys, token-sharded prefill, decode context parallelism, and fused low-latency decode.Model Releases · via vLLM · 3d ago1 min