Open-source AI models in 2026: who leads, and what the 89% Chinese share means

Which open-weight model leads in November 2026?

On Artificial Analysis's ranking of open-weight models, read on November 10, 2026, Moonshot's Kimi K3.1 Pro leads with a score of 46 on its Intelligence Index. The page counts a model as open when its weights can be downloaded.

The Western models on the list are Mistral Large 4 at 26 and Meta's Muse Spark Open at 25.

Rankings differ by method. On Arena's text leaderboard, dated November 9, the top entry is a closed model, and the open-weight models appear below it. A model that leads one list may rank lower on another.

What does the 89% Chinese share mean?

Menlo Ventures' report of November 3, 2026 says Chinese models grew to 89% of open-source spend, led by Kimi at 61%, GLM at 11% and Qwen at 3%, with one other Chinese lab making up most of the rest. Meta's Llama is the most popular Western open model, at 7% of open-source spend.

Three points change how to read it.

• It is a share of a small slice. Menlo says open-source is 4% of spending on models, despite 19% of usage. So 89% is a share of that 4%. The report gives no dollar figure for the slice.
• It is an estimate. Menlo's dollar figures come from a model informed by a survey of nearly 400 U.S. enterprise decision-makers, public filings and third-party reports. The survey was run by Morning Consult in September.
• Its geography is not stated. Menlo's sentence does not say global or U.S. The South China Morning Post describes it as global enterprise spending. Some other outlets call it U.S. enterprise spending. Attribute the wording to the outlet that used it.

Menlo also reports that open-source models handle about one in five tokens but about one in 25 dollars. It puts the average price of an open-source token at $0.13 per million, against $0.70 for closed models on a blended basis. Frontier labs hold 69% of deployment dollars, and Anthropic's share of enterprise API spending rose from 40% to 55%. Menlo is an investor in Anthropic and marks it as a portfolio company, so weigh that when reading its lab figures.

Which releases are real, and which are pending?

Several headline models were announced before their weights could be downloaded. This is the status on November 10.

Tencent's release page says Hy4 places it among the strongest open-source models for the first time, and that a notable gap remains between it and the top closed models. Thinking Machines scores Inkling 2 Small at 31 on the Artificial Analysis index, which is its own report, and says it trails the leading closed models on its alignment tests. It also discloses that its last two training phases, about 354 billion tokens after 12.2 trillion tokens of pretraining, mostly used trajectories sampled from earlier models, including Kimi K3 and its own.

What does open mean, and what are the conditions?

Open-weight means the trained model files can be downloaded. It does not mean open-source in the strict sense, and the licenses differ.

• Permissive. MIT and Apache 2.0 allow broad commercial use. Qwen3.8 27B and Tencent's Hy4 preview use Apache 2.0.
• Custom with conditions. Kimi K3's license is permissive in most uses. A company that runs a model-as-a-service business, and whose group revenue passes $20 million in any 12 months, must first sign a separate agreement with Moonshot. Products with more than 100 million monthly users, or more than $20 million in monthly revenue, must display the Kimi K3 name. Internal use is exempt, and so is use through Moonshot's own products or certified partners.
• Restricted at the top. Qwen's largest model uses a custom license with similar display and service conditions, with a higher revenue threshold for service businesses.

A claim of a revenue share of up to 30% for Kimi K3 appears on a public encyclopedia page. It is not in the license text that Spezia read.

How do open models compare with closed ones?

The makers say they trail. Moonshot's technical report for Kimi K3 says the model trails the leading proprietary models overall. Tencent says a notable gap remains on challenging internal tests. Thinking Machines says Inkling trails closed models on its alignment evaluations.

Government testers say the same about security skills. NIST's Center for AI Standards and Innovation called GLM-5.3 the most cyber-capable open-weight model released so far, and also found it well below current U.S. frontier models on cyber tasks, by about four months on its index. Where they applied, the U.S. models were tested with safeguards off, so the comparison is not like for like. For more on that work, see AI agent security in 2026.

How do you choose an open model?

Five checks help.

1. Are the weights downloadable today? An announcement is not a release.
2. What is the license? Read the conditions on revenue, users and attribution.
3. Who measured the score? A maker's own table is a claim. A third-party index is a measurement of one test.
4. What does it cost to run? Check the price per million tokens and your own hardware.
5. What data goes where? Hosting a model yourself keeps data with you, and using a hosted API does not.