Kimi K3 is the 2.8-trillion-parameter, open-weight model that moved Kimi into the frontier tier in July 2026, with native vision and a 1M-token context. Here are its specs, benchmarks, prices and how it compares with the newer K3.1 Pro.
About 104 billion active for each token.
Up from 256K on K2.6.
Reads images, screenshots and video natively.
Input and output; $0.30 cached input.
| Spec | Kimi K3 |
|---|---|
| Released | 16 July 2026 (apps and API); open weights 27 July 2026 |
| Developer | Moonshot AI |
| Architecture | Mixture-of-Experts with hybrid attention (Kimi Delta Attention plus Gated MLA) |
| Parameters | 2.8 trillion total, about 104 billion active per token |
| Experts | 896 experts, 16 active per token |
| Context window | 1M tokens |
| Input / output | Text, images and video in; text out |
| Reasoning | Always on, with low, high or max effort (default max) |
| Weights format | MXFP4 weights, about 1.4–1.6 TB on Hugging Face |
| Licence | Custom Kimi K3 License (open weights, commercial use allowed) |
| API model ID | kimi-k3 |
| API price | $3 input, $15 output, $0.30 cached input per 1M tokens |
From Moonshot's model card, technical report and API pricing page.
K3 has 896 “expert” sub-networks. For each token, a router switches on just 16 of them, so only about 104B of its 2.8T parameters do the work. That keeps it affordable to run for its size.
It mixes Kimi Delta Attention, a fast linear attention first tested in Kimi Linear, with Gated MLA. This is what makes the 1M-token context practical.
Images and video go into the same model as text, so K3 can read screenshots, charts, documents and video frames directly.
| Benchmark | Kimi K3 | Claude Fable 5 | GPT-5.6 Sol |
|---|---|---|---|
| FrontierSWELong real-world coding projects | 81.2 | 86.6 | 71.3 |
| SWE-MarathonVery long coding tasks | 42.0 | 35.0 | 39.0 |
| DeepSWESoftware engineering agent tasks | 67.5 | 70.0 | 73.0 |
| Terminal-Bench 2.1Command-line tasks | 88.3 | 88.0 | 88.8 |
| BrowseCompFinding hard facts on the web | 91.2 | 88.0 | 90.4 |
| HLE-FullExpert questions, no tools / with tools | 43.5 / 56.0 | 53.3 / 63.0 | 44.5 / 58.0 |
| GPQA DiamondGraduate-level science | 93.5 | 92.6 | 94.1 |
From the Kimi K3 model card, all at max effort. Harnesses differ between models, so treat close results as ties.
K3 led on long coding tasks and web research, but trailed on expert reasoning.
| Test | Kimi K3 | Note |
|---|---|---|
| LMArena Frontend Code Arena | #1 at launch, 1,679 points | Blind human votes on front-end code |
| Artificial Analysis Intelligence Index | 44 (max effort) | K3.1 Pro later scored 55 |
| Vals AI SWE-bench run | 93.4% | Independent run, behind Claude Opus 5 at 97.0% |
| Artificial Analysis speed | About 38–55 tokens/s | Slow for a top model |
| Artificial Analysis cost per task | $2.00 | Writes many tokens per task (about 160M on the index) |
Third-party tests run by the named organisations. Our full take is in the Kimi AI review.
| Kimi K3 | Kimi K3.1 Pro | |
|---|---|---|
| Released | 16 Jul 2026 | 23 Oct 2026 |
| Base model | Kimi K3 | K3 with much more post-training |
| Context window | 1M tokens | 1M tokens |
| API input / output | $3 / $15 | $2 / $10 |
| Cached input | $0.30 | $0.20 |
| Artificial Analysis index | 44 | 55 |
| Kimi Code Bench 3.0 Long-horizon | 55.95% | 61.2% |
| Claim-level hallucination (Moonshot stress test) | 23.4% | 8.3% |
| Faster version | No | K3.1 Pro Highspeed |
Almost everything: it's cheaper per token, stronger at coding and finance, and much less likely to make things up.
You have a tested pipeline that depends on K3's exact behaviour, or you use the k3-256k option in Kimi Code to save quota.
On the API, change the model from kimi-k3 to kimi-k3.1-pro. Nothing else changes.
K3.1 Pro is now the default. K3 powers the extra-long chat mode with up to 1M tokens on the Max and Ultra plans.
Choose k3 (1M context) or k3-256k. The 1M version uses about twice the quota of k3-256k.
Use the model ID kimi-k3 at https://api.moonshot.ai/v1. Batch jobs are billed at 60% of the list price.
Download the open weights from moonshotai/Kimi-K3 on Hugging Face. You'll need a multi-GPU cluster.
Keys and billing: Kimi API guide. Coding setup: Kimi Code guide. Plans: Kimi pricing.
moonshotai/Kimi-K3 on Hugging Face, with the 47-page technical report on Moonshot's GitHub.Artificial Analysis measured about 38–55 tokens per second and very long outputs, so tasks take a while and cost more than the token price suggests.
On Moonshot's own hallucination stress test, K3 made claim-level errors 23.4% of the time, against 8.3% for K3.1 Pro. Check facts.
Heavy demand led to “engine overload” incidents in August 2026, and Kimi paused new paid sign-ups from 20 July until late September.
The weights alone need about 1.4 TB or more of GPU memory, far beyond any single workstation.