Kimi — open frontier models built for long-horizon tasks and deep reasoning
NEW · Kimi K3.1 Pro launched
K3.1 Pro is the current flagship: 1,048,576-token context, $2.00 input / $10.00 output per million tokens (Open Platform list price).
Kimi K3 release
K3 (July 2026): 2.8 trillion parameters, natively multimodal, 1M-token context — built for long-horizon coding, knowledge work and deep reasoning.
Kimi K2.7 Code
Coding specialist: code generation, analysis, refactoring and long-horizon engineering tasks.
This independent guide tracks the Kimi model family — release notes, third-party benchmarks and hands-on coding results — so you can compare before you switch. Or jump straight to TuneRouter to try it.
Benchmarks
Independent evaluation by Artificial Analysis (October 2026): Kimi K3.1 Pro scores 55 on the Intelligence Index, ranking #4 out of 215 models evaluated.
Official blog case study: in GPU kernel optimization (identical sandbox, 24-hour limit, four tasks across NVIDIA Hopper and an alternative-vendor GPGPU), Kimi K3 performed competitively with Claude Fable 5 and substantially outperformed Opus 4.8, GPT-5.6 Sol and GPT-5.5. K3 also completed chip design verification in a single 48-hour autonomous run.
Model lineup
Compiled from official public pages (October 2026). Always defer to official releases.
| Model | Positioning | Context | List price (per M tokens) |
|---|---|---|---|
| kimi-k3.1-proLatest | Current flagship for complex tasks | 1,048,576 | $2.00 in / $10.00 out |
| kimi-k3 | Flagship: 2.8T parameters, natively multimodal | 1,048,576 (K3 series) | See platform |
| kimi-k2.7-code | Coding specialist | See platform | See platform |
| kimi-k2.6 | April 2026 flagship, complex problem solving | See platform | See platform |
Full per-model details on the models page →
Coding, hands-on
What the official tech blog actually ran — not marketing numbers.
GPU kernel optimization
Identical sandbox, up to 24 hours per model, four kernels (AttnRes, KDA, 512-dim MLA) across NVIDIA Hopper and alternative-vendor GPGPU. K3 matched Fable 5 and clearly beat Opus 4.8 and GPT-5.5.
GPU compiler from scratch
K3 built MiniTriton — a Triton-like compiler with its own tile-level IR over MLIR and a PTX codegen pipeline — matching or beating Triton and torch.compile on supported roofline benchmarks.
Chip design in 48h
In one autonomous run, K3 designed, optimized and verified a chip for a nano model of its own architecture, using open-source EDA tools on the Nangate 45nm library.
Trying the API
1. Get a key
Create an account on the Kimi Open Platform (platform.moonshot.ai) and generate an API key.
2. OpenAI-compatible
The platform speaks the OpenAI chat-completions format, so existing SDKs work — point the base URL at the Kimi endpoint and pick a model id (e.g. kimi-k3.1-pro).
3. Or use TuneRouter
TuneRouter routes the same models through one account with unified billing — see the button in the navbar.
FAQ
Who makes Kimi?
Kimi is developed by Moonshot AI. Consumer entry point: kimi.com. Developer platform: platform.moonshot.ai.
Are Kimi models open weights?
The K2 series was released with open weights; licensing differs across generations. Defer to official Moonshot AI announcements.
How do I try Kimi?
Chat at kimi.com or call the API via the Open Platform — or use TuneRouter to access the same models through one account.
What is this site?
An independent third-party information site. Not affiliated with, endorsed by or authorized by Moonshot AI. All data is referenced from public sources; defer to official releases.
Ready to try the Kimi series?
Head to TuneRouter to access these models through a single account.
Go to TuneRouter →