Open a model list today and you’ll see two kinds of names. Claude, GPT and Gemini on one side. DeepSeek, Kimi and GLM on the other. The second group is usually called open weight. The term gets used a lot and explained rarely, so this page explains it.
The weights are the model
A language model is a very long list of numbers. Training sets those numbers, called weights or parameters, and they hold everything the model learned. Kimi K3.1 Pro, one of the largest, has 2.8 trillion of them.
With a closed model, the maker keeps that file. You send a question to the maker’s servers and get an answer back. You rent access and never hold the model.
With an open-weight model, the maker publishes the file. Anyone can download it and run it on their own machines, or pay a hosting company to run it for them.
Open weight isn’t open source
The two terms get mixed up. The Open Source Initiative, which keeps the definition of open source software, wrote one for AI too. Its Open Source AI Definition asks for the freedom to use, study, modify and share a system. That takes three things: the weights, the code that trained the model, and detailed information about the training data.
Open weight covers the first of the three. The OSI’s own page on the term says it refers to the final weights of a trained network, without the training code or the data. You can run the model and adapt it. You can’t rebuild it or see what it learned from.
The license varies too. DeepSeek publishes under MIT, which allows commercial use. Kimi K3, GLM-5.3 and Meta’s Muse Spark Open each come with their maker’s own license, which Artificial Analysis sums up as “commercial use is allowed with restrictions.” Read the license before you build a product on one.
What a published file changes
More than one company can serve it. A closed model is sold by its maker and a few cloud partners. On November 20, Artificial Analysis counted 24 API providers for GLM-5.3 and 21 for DeepSeek V4.1 Flash. Hosts compete on price and speed for the same model. That is a large part of why the cheap tier is cheap, which we covered in the model mix piece.
You choose where it runs. A bank can run the file in its own data center, and nothing leaves the building. A smaller firm can pick a host in the country its contracts require. With a closed model, the maker decides where the servers are.
It can’t be taken back. A host can drop a model from its catalogue. A published file stays published, and another host can pick it up.
It can be adapted. A team with the skills can retrain the file on its own documents. Few small firms will, but the vendors they buy from do.
Buyers have noticed. On Vercel’s AI Gateway, a service many software teams use to reach models, open-weight models carried 71.5% of tokens in September 2026. In December 2025 it was 7%. Vercel notes that its definition of open weight is broader than in its earlier reports, so the two figures aren’t strictly like for like. The direction isn’t in doubt.
Are they as good?
They’re close, and how close depends on who measures.
Epoch AI, a research group, found on May 29 that since January 2026 the best open-weight models have trailed the best closed ones by four months on average. SemiAnalysis, a research firm, went further on August 21. In the free part of its study, it writes that “with each generation, open-source models take half as long to catch up.” In the current generation, by its count, Kimi K2.6 passed Claude Opus 4.5 on its test suite 4.8 months after Opus came out.
Then the tests got harder to study for. On October 27, Artificial Analysis rebuilt its Intelligence Index so that 60% of it rests on private test sets. The gap between the best open-weight model and the best closed one went from 3 points to 12. Kimi K3.1 Pro scored 46, and Claude Opus 5.5 scored 58.
SemiAnalysis adds a caution of its own. A score isn’t the whole story, and its analysts say they still prefer a closed model for their daily work.
So on public tests the two groups are within a few points of each other, and on tests nobody has seen the closed models keep a lead. For drafting, summarizing, sorting and translating, that lead rarely shows. On long, hard jobs it still can.
What you give up
Running the file yourself takes hardware and people. Nobody runs 2.8 trillion parameters on an office laptop. Most firms that use open-weight models rent them from a host, and then the host’s terms apply, not only the license.
Open doesn’t always mean cheap, either. In the same Vercel report, DeepSeek V4.1 Flash handled nearly half of all September tokens for 2.6% of the spend. Kimi K3 was 3.2% of the tokens and drove 47.6% of the rise in open-weight spend. That’s how open-weight models as a group reached 22.6% of spend.
And you still don’t know what the model was trained on.
How we use them at BearingBridge
We offer both kinds. On November 1, our model benchmarks page listed eight open-weight text models from five makers: DeepSeek V4 Pro, V4 Flash and V4.1 Flash, Kimi K3 and K3.1 Pro, GLM 5.3, Muse Spark Open and Mistral Large 4. We checked each one’s status on its Artificial Analysis page on November 20. They sit in the same pickers as Claude, GPT and Gemini.
We don’t run the weights on our own servers. Every call, to an open-weight model or a closed one, goes through one gateway to a hosted endpoint, and lands in the same ledger with its cost.
Three things follow for a customer:
- Your admin decides. Each feature has its own list of allowed models. An organization that wants no open-weight model, or none from a given maker, switches them off.
- You can compare on your own work. The Playground sends one question to up to three models side by side and prints the price under each answer. That tells you more than a leaderboard does.
- The price level is on the page. The two DeepSeek Flash models sit at the lowest of our four price levels. Kimi K3 and K3.1 Pro sit at the highest, beside the most expensive closed models.
We refresh the scores on that page twice a month. The habit we’d suggest to anyone: open-weight models for the bulk work, a closed model where a wrong answer costs money, and a test on real tasks before either becomes the default.
Try it this week
Pick a job your team runs every day, like a product description or a summary of a customer call. Send it to one closed model and one open-weight model. Have the person who knows the work grade both without seeing which is which.
Then ask your vendors which models sit behind their features, and who hosts them.
Last reviewed November 20, 2026. Every source above was opened that day. Model lists, licenses and scores change often. Corrections to hello@bearingbridge.com.







