The I.S. Department Ltd.
ISD Insights · Reference guide · updated monthly

Open-weight model tracker

Open-weight models are AI models that you can download and run yourself: how they score, where they come from, licences and local-run notes. Updated monthly.

Last updated 30 Oct 2026By ISD Insights Team
← Back to ISD Insights
What it means for business
  • Several capable models can now be downloaded and run on your own hardware.
  • Licence terms and each model's origin vary: check both against your own policy.
  • Closed hosted models still score highest on the hardest tasks.

In brief

The highest-scoring Open-weight model: A model whose weights the lab publishes, so you can download it and run it on your own hardware. Glossary models you can download today are DeepSeek V4.1-Flash-1017 and GLM-5.3, tied at 32 on the Artificial Analysis Intelligence Index v5. Meta's Muse Spark Open scores 25. At the small, fast end, Gemma 4 31B, Recursive Flint and Qwen3.8-27B are the cheapest to run locally. Closed models still score higher: 26 points separate the highest-scoring downloadable model from Claude Opus 5.5 (58). We list each model's origin so you can apply your own policy; origin alone says nothing about quality.

Note: Artificial Analysis moved to Index v5 on 27 Oct 2026 and every score dropped. Scores quoted before that date (for example GLM-5.3 at 45) do not compare with the figures below.

Models to know

GroupModelsWhy
Highest scoresGLM-5.3; DeepSeek V4.1-Flash-1017Joint-highest on AA v5 among released weights (32), both China. GLM-5.3 is second on LMArena open (1481). DeepSeek V4.1-Flash-1017 is MIT, with vision and 1M context, and runs in mainline llama.cpp since b11870.
Also strongMiMo-V2.6-ProXiaomi, China. AA 30, MIT, third on LMArena open (1480). Adds image, video and audio input.
Large models under other licencesMuse Spark Open; Reflection Beam 1.0Muse Spark Open (Meta, US) scores 25 on AA v5, under Meta's licence. Reflection Beam 1.0 (Reflection AI, US) is a large Apache-2.0 model; it scores 18 on AA v5 and 1436 on LMArena.
Small and fastGemma 4 31B; Recursive Flint; Qwen3.8-27B; Nemotron 3.5 BoltGemma 4 31B (Google, US) scores 1452 on LMArena, Apache-2.0, multimodal. Flint (Recursive, US) is Apache-2.0 and very fast for coding, but scores below Qwen3.8-27B on accuracy (vendor-reported scores only). Qwen3.8-27B (Alibaba, China) has vision, Apache-2.0. Bolt (NVIDIA, US) is a light agent model (native 2-bit, 22.0 GB).
Watch listKimi K3.1 Pro; Mistral Large 4; Qwen4-27B; Ling 3.1 Flash 1019K3.1 Pro (Moonshot, China) scores 46 on AA v5, 14 points above anything downloadable today, with weights due by 4 Nov. Mistral promised Large 4 weights by end of October; not public on 29 Oct. Qwen4-27B is announced but has no weights. Ling 3.1 Flash 1019 (AA 29, MIT) has only community builds so far.
Less compellingKimi K3; MiniMax-M3.1-Flash; Inkling; Step 5; gpt-oss; GraniteSuperseded (Kimi K3), restrictive licence (MiniMax), weak on independent rankings (Inkling), unverified local support (Step 5), or legacy and niche (gpt-oss, Granite).

Capability

AA = Artificial Analysis Intelligence Index v5 (27 Oct 2026). LMArena = Text, open-source filter (27 Oct 2026).

ModelLab, originReleasedParameters (total / active)AA v5LMArena
DeepSeek V4.1-Flash-1017DeepSeek, China17 Oct 2026552B (763B incl. Engram) / 16B32Not listed (Sep build 1476)
GLM-5.3Z.ai (Zhipu), China14 Aug 2026 (weights 28 Aug)753B / 40B321481
MiMo-V2.6-ProXiaomi, China21-22 Sep 20261.02T / 42B301480
Kimi K3.1 ProMoonshot, ChinaAPI 24 Oct; weights by 4 Nov3T class46 (weights not out)Not listed
Kimi K3Moonshot, China27 Jul 20262.8T / 104BNot in v5 top list1488
Qwen3.8-27BAlibaba, China14 Aug 202627B dense201436
MiniMax-M3.1-FlashMiniMax, China19 Oct 2026428B / 23B29Not listed
Step 5 PreviewStepFun, China20 Sep 2026600B / 27B29Not listed
Ling 3.1 Flash 1019InclusionAI (Ant), China27 Oct 2026560B / 25B29Not listed
Muse Spark OpenMeta, US15 Oct 2026816B / 42B, multimodal25Not listed
Reflection Beam 1.0Reflection AI, US16 Oct 2026501B / 23B, text181436
Recursive FlintRecursive, US21 Oct 202639.8B / 2.5B, textNot ratedNot listed
Gemma 4 31BGoogle, US2 Apr 202631B dense, multimodalNot in v5 top list1452
Nemotron 3.5 Super VLNVIDIA, US15 Oct 2026120B / 12B, vision, 1M141412 (preliminary)
Nemotron 3.5 Bolt 74B-A9BNVIDIA, US15 Oct 202674B / 9BNot ratedNot listed
InklingThinking Machines, US15 Jul 2026975B / 41B, multimodal121442
gpt-oss-120bOpenAI, USAug 2025117B / 5.1BNot in v5 top list1352
Granite 4.2 30BIBM, US25 Aug 202630B denseNot in v5 top list1340
Mistral Large 4Mistral, EUPreview 6 Oct; weights promised end Oct1T / 52B, multimodalNot ratedNot listed
K2 Horizon 375B-A23BIFM (MBZUAI), otherOct 2026375B / 23B18Not listed

Licence and local run

Sizes in GB (10^9 bytes) from the Hugging Face file listings on 29 Oct 2026. As a rule of thumb, a model needs roughly its file size in combined RAM and VRAM: The memory on a graphics card (GPU). Models run fastest when they fit in it. Glossary to run at usable speed.

ModelLicence4-bit download8-bit downloadllama.cpp
DeepSeek V4.1-Flash-1017MIT500.7 GB508.4 GBMainline since b11870 (25 Oct)
GLM-5.3GLM-5.3 licence (MIT-like; very large service providers need review)467.3 GB801.4 GBMainline
MiMo-V2.6-ProMIT554.2 GB (MXFP4)NoneMainline since b11102
Kimi K3.1 ProNot yet publishedNone yetNone yetNot yet verified
Qwen3.8-27BApache-2.017.6 GB29.0 GBMainline
MiniMax-M3.1-FlashMiniMax Community Licence (restrictive)249.4 GB259.1 GBNot verified
Muse Spark OpenMuse licence (attribution; services above $20M need agreement)505.5 GBNoneNeeds a patched build
Reflection Beam 1.0Apache-2.0306.5 GB532.8 GBMainline (b11701)
Recursive FlintApache-2.025.1 GB42.3 GBSupported (vendor)
Gemma 4 31BApache-2.018.8 GB32.6 GBMainline
Nemotron 3.5 Super VLOpenMDW-1.1 (permissive)86.9 GB131.6 GBBuilds published
Nemotron 3.5 Bolt 74B-A9BOpenMDW-1.150.2 GB (native 2-bit 22.0 GB)78.9 GBSupported (NVIDIA)
Mistral Large 4Not yet publishedNoneNoneUnknown

Licences in plain English

  • Permissive (MIT, Apache-2.0, OpenMDW): DeepSeek, MiMo, Qwen3.8-27B, Ling, Beam, Flint, Gemma 4, Inkling, Nemotron. Free for commercial use; keep the licence notices.
  • Conditional, for very large operators only: GLM-5.3 (model-as-a-service above $10B revenue needs review by Z.ai) and Kimi K3 (above $20M needs an agreement). No practical effect for most businesses.
  • Conditional, attribution: Muse Spark Open needs a "Built with Muse Spark Open" notice in products that use it.
  • Restrictive: the MiniMax Community Licence requires a commercial notice to MiniMax and authorisation above $20M. In our view it is best avoided for commercial use.
  • Not yet published: Kimi K3.1 Pro and Mistral Large 4. Read the terms before planning on either.

Governance points

  • Running weights on your own hardware sends no prompts or data to the model's lab, whichever country it is in, which removes the main data-transfer concern about using any lab's hosted service. Licence terms still apply.
  • The US Department of Homeland Security made public on 19 Oct 2026 (US time) an order, issued on 24 Jul, removing DeepSeek products from federal civilian IT systems. The Department of Defense issued its own order on 8 Sep. A Congressional Research Service note says it extends to federal contractors. See our item.
  • Every model reflects choices made in its training, and some reflect the content rules of the jurisdiction it was built in (for example, China-origin models may apply PRC-aligned content filtering). Test any model's outputs on topics that matter to you before relying on them.
  • Hong Kong's PCPD (16 Mar 2026) asks organisations to give agentic AI minimum access rights and to use official builds only.
  • Download from trusted publishers (the labs themselves, unsloth, ggml-org, bartowski, lmstudio-community). Avoid "abliterated" or "uncensored" re-packaged versions.

Sources

← Back to ISD Insights
Contact

Questions about what this means for you?

Talk to us about where AI fits in your business.

[email protected]