PrimaLabs debuts the most intelligent open model. Kimi K3.1.

Why this model matters

Moonshot built K3.1 Pro by post-training Kimi K3 with much more compute, and reports gains across coding, agentic work, 3D creation and financial research.

• Mixture of experts: 2.8T total parameters, 104B active per token.
• One-million-token context window, enough to keep a large codebase or a long agent session in view without trimming history.
• Natively multimodal: text / image input and output on PrimaLabs today.
• Adjustable reasoning: set reasoning_effort to low, high or max (the default).
• Trained for MXFP4: Moonshot used quantization-aware training, so the released MXFP4 weights are the model as trained.
• Kimi K3.1 Pro License, which allows commercial use.

What it's good at

Selected results from Moonshot's model card, at maximum reasoning effort (self-reported, not independently verified):

Next to closed models. On the Artificial Analysis Intelligence Index v5.0.1, an independent composite of nine evaluations across knowledge work, coding and science, Kimi K3.1 Pro scores 46. That is level with GPT-6.1 Sol at medium reasoning, ahead of GPT-6.1 Terra at its xhigh setting, and within two points of Claude Opus 5.5 at medium, against a median of 38 for comparable models:

The top closed models at their highest settings still score higher (Gemini 4 Carbon high 60, Claude Opus 5.5 max 59). For work you would run on those closed models at medium or high settings, K3.1 Pro is an open-weight alternative you can inspect and run where you choose. (Artificial Analysis, accessed 5 November 2026.)

Best fits:

• Coding agents and long sessions: from fixes in real codebases to full-project builds, with the whole session in context.
• Planning and review in multi-model agents: let K3.1 Pro plan and check the work, and hand high-volume steps to a smaller model.
• Document-heavy research: financial research and modeling, reports and office files.
• CAD and 3D: scripts for tools like Blender and CAD software, and parametric parts and layouts.

At its default max effort it reasons at length before answering. For quicker replies, set reasoning_effort to low; for classification, extraction or chat at high volume, a smaller model on PrimaLabs is the better pick (see the table at the end). Moonshot notes its cyber capability trails frontier proprietary models, so it is not the model for security research.

Why run it on PrimaLabs

• The full one-million-token context, not a shortened window.
• Moonshot's native MXFP4 weights, served as trained, with no extra quantization.
• Available the day the weights shipped, with the same API key and endpoint as the rest of our open models.
• Independently benchmarked. On the Artificial Analysis providers page, accessed 6 November 2026, PrimaLabs on Kimi K3.1 Pro is 160 output tokens per second at 10,000 input tokens, 4th of 15. The same page ranks PrimaLabs 4th on time per task at 9.2 minutes, 2nd on time to first answer token at 25.0 seconds, and 1st on cost per task at $2.94.

Get started

Model ID: primalabs-ai/Kimi-K3.1-Pro

curl -sS https://api.primalabs.ai/v1/chat/completions \
  -H "Authorization: Bearer <PRIMALABS_API_KEY>" \
  -H "Content-Type: application/json" \
  -d '{"model":"primalabs-ai/Kimi-K3.1-Pro",
       "reasoning_effort":"high",
       "messages":[{"role":"user","content":"Hello"}]}'

For multi-turn conversations and tool calls, pass each assistant message back exactly as returned, including reasoning_content and tool_calls; Moonshot's model card requires it.

On OpenRouter

Kimi K3.1 Pro is also served by PrimaLabs on OpenRouter (moonshotai/kimi-k3.1-pro). To send your requests to PrimaLabs first, set the provider order:

curl -sS https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer <OPENROUTER_API_KEY>" \
  -H "Content-Type: application/json" \
  -d '{"model":"moonshotai/kimi-k3.1-pro",
       "messages":[{"role":"user","content":"Hello"}],
       "provider":{"order":["primalabs"]}}'

See every model PrimaLabs serves on OpenRouter: openrouter.ai/provider/primalabs.

Pricing

Kimi K3.1 Pro on PrimaLabs, in USD per 1M tokens:

Try it on your workload

Benchmarks tell you where to start; your traffic tells you the answer. Get an API key and run Kimi K3.1 Pro on your own prompts, or talk to an engineer about rate limits for production traffic. Join the community on Discord.

If you need a different model