Why Bells Up AI Users Should Care About Open-Weight Models Like Kimi and GLM

Bells Up AI has added two new model families, Kimi and GLM. These open-weight models were built by AI labs based in China, but when you access these models through Bells Up AI, all of your requests are processed solely in the United States, and the labs that developed these models do not receive what you send or what you receive.

What "open weight" means

Most popular AI models such as Claude (developed by Anthropic) and GPT (developed by OpenAI) are "closed weight" models, which means that the developer does not publicly publish the model "weights," which are used to run the model and produce its responses.

In contrast, an "open weight" model is one where the developer publishes the weights, which anyone can download and run on hardware they control. Some smaller open-weight models can be run on consumer hardware, while larger models can only be run on clusters of computers in data centers.

The Kimi and GLM models available today on Bells Up AI are all open weight models; other open weight models in our model catalog include models from DeepSeek, Alibaba (Qwen), Meta (Llama) and Mistral, which have been in the Bells Up catalog for months.

Contractual privacy provisions when using Kimi and GLM

If you use Kimi or GLM models through Bells Up AI, your data stays inside the United States and the Chinese companies that developed the models do not receive your prompts, your attached documents, or the model output.

We run the newer Kimi and GLM models, Kimi K3, GLM 5.3, and GLM 5.3 Flash, on Fireworks AI. Our agreement with Fireworks AI contractually commits Fireworks to zero data retention, which means that Fireworks will not, unless required by law or in order to provide the service or support to us, "(i) log your Content for human review; or (ii) retain your Content, beyond the time it takes to generate Output and deliver that Output to you." Additionally, Fireworks commits that they will not use your data to train AI models or improve their products. (Bells Up AI makes the same "no training" commitment in our own Terms of Service.) We also run Kimi K3, GLM 5.3, and GLM 5.3 Flash through Fireworks' U.S.-only endpoints, which means that the computers performing the inference are located in the United States.

We access earlier Kimi and GLM releases via Amazon Bedrock. Those models are Kimi K2.5, Kimi K2 Thinking, GLM 5, GLM 4.7, and GLM 4.7 Flash. Our contracts with Amazon similarly prohibit training on your content, and Amazon does not store your prompts or the model's responses after your request is answered, or provide your prompts or the model's responses to the model developer. Bedrock also runs these models inside Amazon's own accounts, in U.S. regions specified in our contract and in our code.

The full data path for each family, including the narrow exceptions to each of these commitments, is set out in our per-family privacy disclosures for Kimi and GLM.

Why lawyers should be open to using open-weight models

Kimi K3 was released in July 2026, and GLM 5.3 and GLM 5.3 Flash were released in September 2026. In our internal testing at Bells Up AI, we have found these models to be useful for both everyday tasks and tasks requiring careful reasoning and analysis. We also believe that the leading closed-weight models, i.e. Anthropic's Claude models and OpenAI's GPT 5.6 models, still produce better responses.

Kimi and GLM models are worth serious consideration because independent evaluators have found them competitive with leading closed-weight models. In its July 2026 assessment (opens in new tab), NIST’s Center for AI Standards and Innovation found that GLM 5.2, the version before GLM 5.3, had overall capabilities similar to OpenAI’s closed-weight GPT-5.2, released in December 2025. In Arena’s blind-comparison Legal & Government leaderboard (opens in new tab), Kimi K3 Max ranked first on August 12, 2026, ahead of closed-weight models in that category.

Closed-weight models remain under the developer’s control. A developer can severely restrict availability, as Anthropic has done with its Claude Mythos model, withdraw a model from general availability, or change its pricing. Open-weight models can be hosted by multiple providers, reducing dependence on a single developer and making continued access more likely.

We believe that lawyers should also be aware of open-weight models because, as they become more widely used, you will encounter their output. Opposing counsel may use an open-weight model in their own work, or may use an open-weight model as part of a product that they buy. For those reasons alone, we believe that running the same model yourself and comparing its answer with another model can help you in your own practice.

In August 2026, Harvey announced (opens in new tab) that its new legal model, Harvey Tenet, has been built on Kimi K3. To conduct post-training on Kimi K3 to create Harvey Tenet, Harvey used Fireworks.ai, the same platform that Bells Up AI is using to access Kimi K3.

Harvey reported that Tenet performs better than the base Kimi K3, on Harvey's internal legal agent benchmark, and coverage of the announcement has relayed claims that it outperforms frontier models from Anthropic and OpenAI. We note that Harvey's performance claims rely on Harvey's own benchmarks, and that no independent evaluator has confirmed them.

Harvey has not released Tenet's weights, and selecting Kimi K3 in Bells Up AI does not give you access to Harvey Tenet. Instead, Harvey took the open Kimi K3 weights and conducted additional model training, which is something Harvey could do precisely because the weights are open. What you get through Bells Up AI is the same underlying model that Harvey started from for Harvey Tenet, on the same infrastructure company Harvey used, for $99 per lawyer per month.

Export controls and open US investigations

As of September 10, 2026, Bells Up AI has not identified a U.S. federal or state restriction that prohibits a private law firm from using Kimi or GLM through Bells Up AI. The U.S. government has imposed restrictions directly on certain Chinese AI companies and has opened inquiries into others. Those actions do not create a general ban on private or commercial use of Kimi or GLM through Bells Up AI. They may, however, restrict use for a private firm performing work for a government client. Texas prohibits listed technologies on state-owned devices and networks, and imposes additional restrictions on personal devices used to conduct state business. Firms should assess the requirements of each government engagement.

Z.ai, the developer of GLM, has appeared on the U.S. Commerce Department's Entity List since January 2025. That listing restricts the U.S. technology that can be shipped to Z.ai. It does not purport to prohibit anyone from using Z.ai's published model weights.

In July 2026, U.S. officials alleged that Moonshot used another developer's model outputs and export-restricted hardware to build Kimi K3. Moonshot is not currently on the Entity List. The Bureau of Industry and Security opened an investigation, and the Treasury Secretary said publicly on July 22, 2026 that sanctions and an Entity List designation were on the table. No charges, sanctions, or export-control penalties have been imposed.

On September 8, 2026, the NSA, CISA, and the FBI issued a joint advisory, AA26-251A, naming six China-based AI companies, including Moonshot AI and Z.ai, accusing the companies of extracting training data from U.S. frontier models. The advisory states that Moonshot AI "extracted significant Claude Fable 5 data to train its Kimi-K3 model and GPT-4o data to train its Kimi-K2 model," and that Z.ai distilled GPT-5.5 and Claude Opus 4.8 data. Its recommendations are directed at U.S. AI developers, on detecting and responding to this practice. It imposes no restriction or obligation on companies or individuals that use these models. It concerns how the models were trained, not what happens to your prompts.

The House Select Committee on the CCP and the House Homeland Security Committee also have an open joint inquiry into U.S. companies' use of Chinese open-weight models. The committees have demanded information from major US companies, covering data security, supply-chain concerns, and procurement.

Read the disclosure, then decide

Each model family has a full privacy disclosure inside Bells Up, available within two clicks of login. The Kimi and GLM disclosures identify each party that handles your data, quote the relevant contract sections, and describe the export-control situation, the September advisory, and the unresolved research concerning content shaping. Read them before deciding whether to use either family.

Your firm's administrator can disable either family for every firm user. Individual users can also hide a model family from their own accounts.

We hope you will take this opportunity to try out Kimi K3, GLM 5.3, GLM 5.3 Flash, and the other models in our catalog, and assess for yourself whether they may be suitable for any of your work. As a starting point, try Kimi K3 on your next task and use our "Compare" feature to compare its output side-by-side with Anthropic's Claude or OpenAI's GPT.

If you have any questions or comments, or if you find something we should know, tell us at support@bellsup.ai. We love hearing from our customers, and a human being reads all of your messages.