Skip to content
All writing

The best open-weight AI models in November 2026

Aarav Pundir · FounderOpen-weight models5 min

Kimi K3.1 Pro, DeepSeek V4.1 and GLM-5.3 lead open-weight AI in November 2026. How far they trail Claude, GPT and Gemini on independent tests.

The most capable open-weight model in November 2026 is Moonshot AI's Kimi K3.1 Pro. On the Artificial Analysis Intelligence Index, read on 19 November, it scores 46. The best closed model, Anthropic's Claude Fable 5.2 at maximum effort, scores 65. The next open models are DeepSeek V4.1 Flash at 33 and Z.ai's GLM-5.3 at 32. Measured over time, open models trail the closed frontier by roughly four months. Almost every leading open model now comes from a Chinese lab. Meta's best open model, Muse Spark Open, is the strongest American one.

What does open-weight mean?

The trained model file is published, so you can download it and run it on your own hardware or with any host. It is different from open source in the full sense. The training data and code are usually not released, and the licence can attach conditions. We cover those conditions in should you run coding agents on open-weight models.

What is the best open-weight model right now?

Kimi K3.1 Pro, by a clear margin on independent tests. Moonshot launched it through its API on 23 October and published the weights on Hugging Face in early November. It is a mixture-of-experts model with 2.8 trillion parameters in total, of which 104 billion are active for each token. Its context window is about one million tokens.

It is strongest at agent work. Moonshot reports 65.7 on Terminal-Bench 4.0, a test of tasks done in a real terminal. Artificial Analysis ran the test itself and measured 67%, the highest score of any model on the day it launched.

Which other open models matter?

These are the open-weight releases most teams are choosing between. Sizes are total parameters and, for mixture-of-experts models, the parameters active per token. All details come from each model's own page.

ModelLabWeights outSizeLicence
Kimi K3.1 ProMoonshot AINovember 20262.8T, 104B activeCustom
Qwen3.8-MaxAlibabaAugust 20262.4T, 95B activeCustom
DeepSeek V4-ProDeepSeek20261.6T, 49B activeMIT
DeepSeek V4.1 FlashDeepSeek10 September 2026552B, 8B to 16B activeMIT
GLM-5.3Z.ai25 August 2026753BCustom
GLM-5.2Z.ai16 June 2026753BMIT
Muse Spark OpenMeta15 October 2026816B, 42B activeCustom
Mistral Large 4Mistral AIOctober 2026about 1TModified MIT
Gemma 4 31BGoogle2 April 202631BApache 2.0
Nemotron 3.5 Super VLNvidia15 October 2026120BOpen weights and data

Two names are easy to confuse. Alibaba announced Qwen4-Max on 16 November, but it is available through Alibaba's API only. Alibaba says smaller open Qwen4 models are coming. Qwen3.8-Max, from the summer, is the one with open weights.

How far behind closed models are open models?

About four months, or more, depending on how you measure.

In months. Epoch AI tracks the gap with its own capabilities index. It found that since January 2026 open models have trailed the closed frontier by an average of four months, up from about three months in October 2025. Epoch says the method may understate the gap. The analyst Nathan Lambert puts the best Chinese open models two to five months behind the closed US frontier, and American open models six to nine months behind.

In a government test. The US Center for AI Standards and Innovation (CAISI), part of NIST, now evaluates major open releases. It judged GLM-5.2 probably the most capable open-weight model when it came out in June, with overall ability similar to OpenAI's GPT-5.2 from December 2025. It found GLM-5.3 about four months behind the US frontier on cyber tasks, and DeepSeek V4 about eight months behind.

In index points. That is where the answer moves most, which is the next question.

Why does the size of the gap depend on the test?

Because public benchmarks flatter every model that has seen them, and open models are usually tested on public benchmarks.

Artificial Analysis showed this clearly. Under version 4.3 of its index, the best open model sat 3 points behind the best closed one. On 27 October it launched version 5, which takes 60% of its score from private or held-out tests. The gap widened to 12 points overnight. With Fable 5.2 added in November, it is 19.

CAISI found the same pattern. It built its own uncontaminated coding test, PortBench, which no model could have trained on. GLM-5.2 scored 42% on it, against 60% for Claude Opus 4.6 and 78% for GPT-5.5. On the public SWE-bench it scored 75%, much closer to the leaders.

The practical lesson: a vendor's launch table tells you the most a model can score, not what it will do with your code.

Why do Chinese labs lead open models?

Because they kept releasing while most American labs stopped.

Hugging Face reported in March that Chinese models made up 41% of downloads on its platform over the previous year, and that China had passed the US in monthly downloads. Alibaba's Qwen family alone had about 2.05 billion downloads from January to August 2026. Developers built 151,448 derived models on Qwen, against 82,506 on Google's models.

The same shift shows up in real use. On OpenRouter, a service that routes requests to many models, open-weight models carried 26.8% of tokens in October 2025. They peaked at 77.5% in July 2026. In the week to 18 November, the most used model on OpenRouter was DeepSeek V4.1 Flash, with 62.1 trillion tokens. Money tells a different story. Closed models still took 73% of spend in September, because they cost far more per token.

Did Meta stop releasing open models?

No, but its best model is no longer open. Meta launched Muse Spark in April 2026 without open weights. It has since released a small open model, Muse Glimmer, in August, and Muse Spark Open on 15 October. Meta calls it the most capable American open-weight model. Its licence limits commercial use.

Elsewhere in the US, Google's Gemma 4 moved to the Apache 2.0 licence in April, Nvidia released Nemotron 3.5 in October with its training data, and the startup Reflection announced its Beam model in October. OpenAI's Hugging Face page shows no successor to its 2025 gpt-oss models this year.

How should you choose an open model?

Test it on your own work. Every number in a launch post was run by the vendor, on a harness the vendor chose, and the independent results above show how far those can drift.

  • Shortlist by independent scores, such as Artificial Analysis, CAISI or Epoch, not by the vendor's table.
  • Read the licence file, not the summary. Several labs moved from MIT to custom licences this year.
  • Run a fixed set of your own tasks with the same instructions on each candidate, and judge the results with the same written test.

That last step is where a model choice becomes safe to change later. If every task carries a definition of done and a proof of work, you can compare models on results you have checked rather than on claims. See what a proof of work contains.