ScaledThought

Changelog

The open-weight release log.

Every model release we track, newest first. Dates are when the lab published the weights. Scores are the Artificial Analysis Intelligence Index where available.

September 2026

  1. Sep 22, 2026MiMo-V2.6-ProNewXiaomiXiaomi's flagship omnimodal MoE model, and the top-ranked open-weight model on the Artificial Analysis Intelligence Index.1.0T / 42B active · 1M context · MIT46AA Index
  2. Sep 10, 2026DeepSeek-V4.1-FlashNewDeepSeekDeepSeek's efficient multimodal update to V4 Flash, MIT-licensed with a 1M-token context window.552B / 16B active · 1M context · MIT39AA Index
  3. Sep 3, 2026K2 Horizon 375BNewIFMFlagship of the Institute of Foundation Models' fully open K2 Horizon family, released with weights, code and training data.375B / 23B active · 512K context · Apache 2.031AA Index

August 2026

  1. Aug 29, 2026GLM-5.3Z.aiZ.ai's flagship agentic and coding model, with dynamic sparse attention and a 1M-token context window.753B / 40B active · 1M context · GLM-5.3 License (custom)45AA Index
  2. Aug 28, 2026Hy4 previewTencentTencent's largest open-weight MoE yet, aimed at coding, office and scientific workflows.770B / 49B active · 1M context · Apache 2.0
  3. Aug 26, 2026GLM-5.3-FlashZ.aiThe smaller, cheaper multimodal sibling of GLM-5.3, built for high-throughput agent work.320B / 18B active · 1M context · MIT (reported)42AA Index
  4. Aug 14, 2026Qwen3.8-27BAlibabaDense, natively multimodal workhorse that fits on a single GPU. Passed 1M downloads in two days.27.8B · 256K context · Apache 2.034AA Index
  5. Aug 13, 2026DeepSeek V4 ProDeepSeekDeepSeek's largest current model. Previewed in April, with open weights at general availability in August.1.6T / 49B active · 1M context · MIT36AA Index
  6. Aug 12, 2026Qwen3.8-2.4TAlibabaThe open checkpoint behind Qwen3.8-Max, the first time Alibaba has released a Max-class model's weights.2.4T / 95B active · 961K context · Apache 2.040AA Index
  7. Aug 10, 2026Muse GlimmerMetaMeta's first open-weight release since Llama 4: a dense agent model distilled from Muse Spark, sized for one 24GB GPU.30B · Apache 2.017AA Index

July 2026

  1. Jul 27, 2026Kimi K3Moonshot AIMoonshot's 2.8T-parameter model, with a new hybrid linear-attention architecture and native vision.2.8T / 104B active · 1M context · Kimi K3 License (custom)44AA Index
  2. Jul 15, 2026InklingThinking MachinesThinking Machines Lab's first open model, trained on 45T multimodal tokens and built to be fine-tuned.975B / 41B active · 1M context · Apache 2.025AA Index

June 2026

  1. Jun 4, 2026Nemotron 3 UltraNVIDIANVIDIA's largest Nemotron 3: a hybrid Mamba-Transformer MoE tuned for agents and tool calling.550B / 55B active · OpenMDW-1.123AA Index
  2. Jun 1, 2026MiniMax-M3MiniMaxMiniMax's natively multimodal coding and agent model, strong on real-world software engineering benchmarks.428B / 23B active · 1M context · MiniMax Community License29AA Index

April 2026

  1. Apr 2, 2026Gemma 4 31BGoogle DeepMindThe largest Gemma 4 model. The family moved from Google's custom license to full Apache 2.0.31B · 256K context · Apache 2.019AA Index

March 2026

  1. Mar 16, 2026Mistral Small 4Mistral AIMerges Mistral's Instruct, reasoning and coding lines into one fast multimodal MoE.119B / 6B active · 256K context · Apache 2.0

December 2025

  1. Dec 1, 2025Mistral Large 3Mistral AIMistral's largest Apache 2.0 model, and its first fully permissive flagship.675B · 256K context · Apache 2.0

November 2025

  1. Nov 20, 2025OLMo 3 32BAi2Ai2's fully open reference model: weights, training data, code and every intermediate checkpoint.32B · Apache 2.0

August 2025

  1. Aug 5, 2025gpt-oss-120bOpenAIOpenAI's first open-weight model since GPT-2. Runs on a single 80GB GPU and is still a widely used workhorse.117B / 5.1B active · 128K context · Apache 2.0

June 2025

  1. Jun 5, 2025Qwen3-Embedding-8BAlibabaTopped the MTEB multilingual leaderboard at release, and still the standard open embedding family.8B · 32K context · Apache 2.0

April 2025

  1. Apr 5, 2025Llama 4 MaverickMetaMeta's 128-expert natively multimodal MoE, still widely deployed as a dependable workhorse.400B / 17B active · 1M context · Llama 4 Community License

Watch: npx -y scaledthought@latest models --live, or /llms.txt

Claude Codeterminal↗Claude Desktopcode tab↗Codexapp↗

Opens your agent with the prompt typed. You press Enter.