Kimi K3
Open weights can be deployed locally (large GPU required).
01
Snapshot
Overall
77
rank #8 of 22
Coding
71
rank #11 of 22
Multimodal
81
rank #10 of 22
Input / Output
$3 / $15
USD per 1M tokens
Context
256K
usable recall 99
Speed
~45 tok/s
TTFT 0.7s
Aggregated from public sources and independently weighted; methodology on the Terms page. Scores within 3 points are treated as statistically tied.
02
Specs & pricing
Every tracked field. List API prices in USD per 1M tokens; verify current vendor pricing before purchase.
| Vendor | Moonshot AI (CN) |
| Released / current | 2026.06 |
| Licensing | Open weights |
| Context window | 256K |
| Max output | 64K |
| Effective-context score | 99 |
| Input / output price | $3 / $15 per 1M tokens |
| Cache discount | Context-cache discount (rate not published) |
| Free tier | Kimi App free; free quota on the open platform |
| Speed / TTFT | ~45 tok/s / 0.7s |
| Function calling | 80 |
| Refusal rate | ~6% |
| English / Chinese | 75 / 92 |
| Modalities | text, image |
| Fine-tuning | No |
| Private deployment | Yes (enterprise) |
| SOC2 / no-train | no / yes |
03
Strengths & watch-outs
Strengths
- Best-in-class long text
- Open-source and self-hostable
- Strong codebase understanding
- Fair price
Watch-outs
- Average multimodal
- Mid reasoning depth
04
Where it ranks among 22 models
Overall 77/100 (#8), coding 71 (#11), multimodal 81 (#10). See the full boards and side-by-side compare on the leaderboard.
05
Pricing in practice
Illustrative monthly bill for 100M input + 30M output tokens: $750 at list. Your mix and cache-hit ratio change this.
Illustrative model using public list rates before any enterprise agreement; cache discount: Context-cache discount (rate not published).
06
Frequently asked questions
How much does Kimi K3 cost?
List API pricing is $3 input and $15 output per 1M tokens. Context-cache discount (rate not published). Kimi App free; free quota on the open platform
What is Kimi K3's context window, and is it usable end to end?
It advertises a 256K window with 64K max output; its effective-context score is 99/100, which is the better predictor of whether long-document details are actually retained.
Is Kimi K3 good for coding?
Its coding aggregate is 71/100, rank #11 of 22. It is a strong choice for most product engineering.
Can I self-host Kimi K3?
Yes — it ships open weights and supports private deployment, so you can run it on your own infrastructure for data control; budget for GPU and operations.
How strong is it in Chinese and on multimodal inputs?
Chinese score is 92 versus English 75; multimodal aggregate is 81/100 (rank #10), accepting text, image.