Kimi K3.1 Pro: A 3T-Class Open-Weight Model With 1M Context
Moonshot shipped the first model in its K3.1 series on 24 October. Kimi K3.1 Pro is a 3T-class open-weight flagship with a 1-million-token context window, built as a post-trained version of Kimi K3 with substantially larger post-training compute and joint reinforcement learning across hundreds of downstream domains. A Highspeed variant runs 2-3x faster, and per Moonshot the full weights land by 4 November
Tool Interface
Interactive tool will be available soon
Features
- ✓ A 1-million-token context window with flat pay-as-you-go pricing and no tiering by context length
- ✓ A 3T-class open-weight flagship that Moonshot says delivers substantial gains in coding, 3D creation, financial research and other agentic capabilities
- ✓ Coding spans small fixes in real codebases to full-project builds, working across Unity, Unreal Engine and Godot alongside web search and 3D generation, with stronger composition and motion in Three.js scenes
- ✓ On the 3D side it does parametric modelling: Moonshot says it writes and runs scripts for Blender and CAD tools directly, without MCP connectors, to build assets, modify parts and adjust geometry, materials and scene layout
- ✓ Moonshot reports state-of-the-art results on OfficeQA Pro v2 and SpreadsheetBench v2 for financial work, and on de-identified conversations where users had flagged factual errors, claim-level hallucination fell from 23.35% on K3 to 6.85%
- ✓ For the open-weight release Moonshot removed cyber-related data from both mid-training and RL while keeping general coding ability; a kimi-k3.1-pro-highspeed variant is listed at 2-3x the speed
How to Use
- Get an API key at platform.kimi.ai, point base_url at https://api.moonshot.ai/v1 and use the model ID kimi-k3.1-pro. It is a flagship model unlocked after a top-up of at least $1, and your cumulative top-up also sets concurrency and rate limits
- Watch the fixed parameters: max_completion_tokens defaults to 131072 and can go up to 1048576, while temperature=1.0, top_p=0.95, n=1 and both penalties are fixed, so omit them from requests
- In multi-turn and tool-calling flows, return the complete assistant message unchanged. For images, do not use public URLs: send base64 or ms:// and make content an array of objects
- It always thinks and you cannot turn that off, so lower reasoning_effort if latency hurts. Moonshot also notes web search is being updated and is not recommended for production workflows in the near term
FAQ
When did Kimi K3.1 Pro launch, and are the weights out?
Moonshot AI introduced it on 24 October 2026 as the first model in the K3.1 series. The docs state that full model weights will be released by 4 November, and the API and apps are available now. Docs: https://platform.kimi.ai/docs/guide/kimi-k3-1-pro-quickstart
How does it relate to Kimi K3?
Per Moonshot, K3.1 Pro is a post-trained version of Kimi K3 with substantially larger post-training compute. To make large-scale joint RL across hundreds of downstream domains stable, they reworked the RL framework, long-context inference, training efficiency, sandbox infrastructure, rollout scheduling and a shared platform for environment quality checks.
How do context and billing work?
It offers a 1M-token context with flat pay-as-you-go pricing and no tiering by context length. Billable items are cache writes (with separate 5-minute and 1-hour TTL tiers), input (with separate rates for cache hits and misses) and output, all at uniform per-token prices. max_completion_tokens defaults to 131072 and can be set up to 1048576.
What are the hard constraints?
temperature=1.0, top_p=0.95, n=1, presence_penalty=0 and frequency_penalty=0 are all fixed and should be omitted. Vision input does not accept public image URLs, so use base64 or ms://. Chain-of-thought cannot be disabled, only reduced via reasoning_effort. Moonshot also says web search is being updated and is not recommended for production workflows for now.
What did Moonshot do for the open-weight release?
To limit misuse, Moonshot says it identified and removed cyber-related data from both the mid-training corpus and the RL environment mix while keeping the general-purpose coding tasks, then evaluated vulnerability research on its in-house xdev-bench-hard, which reproduces recently disclosed vulnerabilities in software such as the Linux kernel, Node.js, JavaScriptCore, Exim and VirGL.