The gap between open and closed artificial intelligence models narrowed again this week. Moonshot released the weights for Kimi K3.1 Pro, and Mistral did the same for Large 4. Both pitch finance as a core use case.
Moonshot, Mistral and Tencent Open Their Flagships
- Moonshot AI released the weights for Kimi K3.1 Pro. Weights are the numbers a model learns during training that decide how it responds to any input. Releasing them lets anyone download the model and run it on their own servers instead of calling a vendor’s API. The Beijing-based lab says its 2.8-trillion-parameter model brings major gains in coding, 3D design and financial research. Parameters are those same learned numbers counted up; more usually means more capable and more expensive to run. On a finance agent test, it scored 58.6 against 45.0 for OpenAI’s GPT-6 Astra. It has a 1-million-token context window, enough to hold a full deal file in one pass. Tokens are the small chunks of text AI companies use to meter usage.
- Mistral AI released the weights for Mistral Large 4 on Oct. 29. The Paris-based company nicknamed the trillion-parameter model “Le Chonk” and aimed it at cybersecurity. Mistral says companies can run it on their own machines to find and fix vulnerabilities, with no vendor able to revoke access. It ships under a modified MIT license. The MIT license is one of the most permissive in software: anyone can use, change and sell products built on the model with almost no conditions. Mistral’s version adds one, requiring large cloud providers that resell the model as a service to sign a separate agreement.
- Tencent launched Hy4 on Wednesday (Nov. 4) and said open weights will follow in the coming days. Tencent trained it on thousands of financial workspaces covering fund research, securities analysis and quantitative strategies. In a blind evaluation by 213 financial professionals, it ranked first among open models, ahead of DeepSeek and Kimi. Tencent says it cut the hallucination rate from 12.95% to 7.2%. Pricing is about $0.90 per million input tokens and $2.70 per million output tokens.
Z.ai and Nvidia Build for Legal Review and Call Transcription
- Z.ai opened its GLM-Legal model to public API access on Monday (Nov. 2), after an invite-only beta that began Oct. 12. The Chinese lab, formerly Zhipu AI, built it for contract review, evidence analysis and cross-border compliance. Its demos include a contract dispute timeline with every fact cited to a source document.
- Nvidia released Nemotron 3.5 Transcribe on Oct. 28, an open-weight speech recognition model. It transcribes 40 languages, labels up to eight speakers and translates 31 languages to or from English. Nvidia says it ranks first on Hugging Face’s open speech recognition leaderboard. It runs on a single GPU, so a bank could transcribe customer calls in-house without sending audio to a cloud provider.
Google and Artificial Analysis Change What Gets Measured
- Google DeepMind rolled out Genie 4 to Google AI Ultra subscribers this week. Genie 4 is a world model: It takes a text prompt or a single image and generates a 3D environment the user can walk through, with sound, rendered live as they move. Objects stay where they were left and the scene holds together over time, which earlier versions could not manage. Google says the significance is in training. Robots and software agents can practice in thousands of generated worlds before they touch the real one, at a fraction of the cost of building physical test setups or collecting real-world data.
- Artificial Analysis launched version 5 of its Intelligence Index on Oct. 29, the independent scorecard most labs cite when claiming the top spot. For the first time, more than half the score comes from private benchmarks that labs cannot train on in advance. Knowledge work, coding, general intelligence and science now count equally. Google’s Gemini 4 Carbon leads at 59.
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.
