Moonshot Launches Kimi K3.1 Pro With Million-Token Context

Moonshot AI has launched Kimi K3.1 Pro, a new model aimed at coding agents, financial research and 3D creation. The Chinese company says it has a one-million-token context window and offers a Highspeed version that responds two to three times faster. Both versions are available through Kimi’s apps and API now, while the full model weights are promised by 4 November. That last distinction matters. K3.1 Pro is being described as open-weight, but its weights are not yet available to download.

In its technical announcement, Moonshot presents the release as a post-trained version of Kimi K3 rather than an entirely new base model. It says additional reinforcement learning improved long-running tasks in software development, office work and research. For developers, the practical question is whether those gains hold up in their own repositories and workflows, not just in charts selected for a launch.

Moonshot reports a 65.7% score for K3.1 Pro on Terminal-Bench 4.0, a test of coding agents working through terminal tasks. Its published comparison puts the model ahead of GPT-6 Astra at 58.2% and just behind Claude Opus 5.5 at 66.4%. These are vendor-reported comparisons with specific evaluation settings. They suggest a competitive coding model, but they are not a guarantee that it will outperform either rival on every project.

The company is also pushing beyond chat and code completion. It says K3.1 Pro can build projects using Unity, Unreal Engine and Godot, and has improved its ability to generate and refine CAD and Blender assets. A new 3D Make feature is live in the Kimi app, allowing users to preview and edit models, then export them for 3D printing or other uses. The launch adds a tangible product demonstration to Moonshot’s benchmark claims, although readers should still judge the quality of generated files in the tools they actually use.

Financial research is another focus. Moonshot says K3.1 Pro scores 64.4 on OfficeQA Pro v2 and performs better than K3 on internal tests of modelling and source-grounded research. It also says the claim-level hallucination rate on its difficult internal test fell from 23.4% to 8.3%. The company cautions that this stress test is designed to provoke mistakes and should not be read as an everyday error rate. Even an improved model can make consequential errors in an investment spreadsheet, so human review and source checks remain essential.

K3.1 Pro is already accessible on Kimi.ai, the Kimi mobile apps, Kimi Work, Kimi Code and the Kimi API. Developers can choose the standard or Highspeed API model. Full weights are due by 4 November, according to Moonshot, so organisations considering self-hosting still need to wait for the actual release and its licence terms. Open weights would offer more control over deployment, but a model in this size class also requires substantial infrastructure.

Moonshot has made a notable safety trade-off for that planned release. It says it removed cyber-related material from parts of training after observing stronger vulnerability-research ability as post-training scaled. On its internal cyber test, the final model scored 81.0% against 90.5% for a checkpoint without those risk-reduction measures. That is the company’s own evaluation, but it makes the intended compromise explicit. The lab wants a useful open-weight coding system without releasing the strongest offensive cyber capability its training pipeline could produce.

For the wider AI market, K3.1 Pro extends a pattern TechBooky has followed since Kimi K3 emerged as a serious coding contender. Customers now have more choices among capable models, and those choices come with different prices, access conditions and data-governance questions. Moonshot’s new release deserves testing because it is available, not because one score settles the race. The next milestones are independent evaluations and the promised weight release.