China AI Hub
Models / Kimi K3
Kimi K3
Kimi K-series family · Provider: moonshot-ai · Status: active · Released: 2026-07-16
Kimi K3 is a Kimi K-series model developed by moonshot-ai, a Chinese AI company headquartered in 13F, Building 1, JD Technology Building, 76 Zhichun Road, Haidian District, Beijing, China, released 2026-07-16 with a 1,048,576-token context window and open weights.
| Key fact | Value |
|---|---|
| Model ID | kimi-k3 |
| Architecture | MoE: 2.8T total / 104B activated; 93 layers (1 dense); 896 experts (16 selected + 2 shared per token); 69 KDA + 24 Gated MLA layers; hidden dim 7168; SiTU-GLU; MoonViT-V2 vision encoder (401M); MXFP4 weights / MXFP8 activations |
| Context window | 1,048,576 tokens |
| Max output | 1,048,576 tokens |
| Open weights | Yes |
| License | Kimi K3 License (permissive MIT-style, but Model-as-a-Service operators with >$20M aggregate revenue over any 12 months must sign a separate agreement; products with >100M MAU or >$20M monthly revenue must display 'Kimi K3' in the UI) |
| Self-hosting | Yes |
| API available | Yes |
| API pricing | $3 input / $15 output per 1M tokens (USD) · provider pricing page |
| Cloud providers | Moonshot AI Platform |
Capabilities
| Capability | Supported |
|---|---|
| Reasoning | Yes |
| Coding | Yes |
| Math | Unknown |
| Chinese | Unknown |
| English | Unknown |
| Multilingual | Unknown |
| Vision | Yes |
| Audio | Unknown |
| Video | Yes |
| Tool calling | Yes |
| Function calling | Unknown |
| Structured output | Yes |
| Agent capability | Yes |
| RAG | Unknown |
| Computer use | Unknown |
Benchmark results
| Benchmark | Version | Score | Metric | Date | Source type | Source |
|---|---|---|---|---|---|---|
| GPQA Diamond | — | 93.5 | accuracy | 2026-07 | vendor_reported | link |
| HLE-Full | — | 43.5 (56.0 with tools) | accuracy | 2026-07 | vendor_reported | link |
| DeepSWE | — | 67.5 | accuracy | 2026-07 | vendor_reported | link |
| Terminal-Bench 2.1 | — | 88.3 | accuracy | 2026-07 | vendor_reported | link |
| MMMU-Pro | — | 81.6 | accuracy | 2026-07 | vendor_reported | link |
| Video-MME (with subtitles) | — | 90 | accuracy | 2026-07 | vendor_reported | link |
Benchmark scores are single data points, not universal rankings. Vendor-reported scores are labeled as such.
Known limitations
- Temperature fixed at 1.0 and top_p at 0.95 - cannot be modified
- API access requires a minimum $1 top-up
- Modality documentation inconsistent: architecture table says Text+Image, while the README, launch blog and API guide also show video input
- Benchmarks vendor-reported; some comparison scores cited from Artificial Analysis
Kimi K3 is Moonshot AI’s flagship open-weight model (“Open Frontier Weights”, released 2026-07-16): a 2.8T-parameter MoE with 104B activated parameters, a 1M-token context window, native visual understanding, and always-on thinking with reasoning_effort low/high/max (default max). It targets software engineering, knowledge work and deep reasoning.
The API model kimi-k3 (min $1 top-up) prices at $3.00 input / $15.00 output per 1M tokens with
automatic prefix caching ($0.30 cached input); max_completion_tokens defaults to 131,072 and can be set
up to 1,048,576. Temperature (1.0) and top_p (0.95) are fixed. Full weights are on Hugging Face and
ModelScope under the Kimi K3 License.
Provider
Pricing
Benchmarks with results for this model
Agents built on this model
API
Data interpretation
| Field | Value | Evidence type |
|---|---|---|
| Context window | 1,048,576 tokens | Official |
| Architecture | MoE: 2.8T total / 104B activated; 93 layers (1 dense); 896 experts (16 selected + 2 shared per token); 69 KDA + 24 Gated MLA layers; hidden dim 7168; SiTU-GLU; MoonViT-V2 vision encoder (401M); MXFP4 weights / MXFP8 activations | Official |
| Open weights | Yes | Official |
| License | Kimi K3 License (permissive MIT-style, but Model-as-a-Service operators with >$20M aggregate revenue over any 12 months must sign a separate agreement; products with >100M MAU or >$20M monthly revenue must display 'Kimi K3' in the UI) | Official |
| API pricing | $3 / $15 per 1M tokens (USD) | Official |
| GPQA Diamond | 93.5 accuracy | Vendor-reported |
| HLE-Full | 43.5 (56.0 with tools) accuracy | Vendor-reported |
| DeepSWE | 67.5 accuracy | Vendor-reported |
| Terminal-Bench 2.1 | 88.3 accuracy | Vendor-reported |
| MMMU-Pro | 81.6 accuracy | Vendor-reported |
| Video-MME (with subtitles) | 90 accuracy | Vendor-reported |
| Family position | 1 model in the Kimi K-series family | China AI Hub analysis |
Evidence types: Official = vendor documentation, pricing pages or model cards. Vendor-reported = benchmark scores published by the vendor. China AI Hub analysis = derived from the database itself. See the sourcing policy.
Sources
Confidence and source hierarchy per the sourcing policy. Facts change; check the source before relying on this page.
What is the context window of Kimi K3?
Kimi K3 has a 1,048,576-token context window and a maximum output of 1,048,576 tokens.
Is Kimi K3 open weight?
Yes — Kimi K3 weights are openly available under the Kimi K3 License (permissive MIT-style, but Model-as-a-Service operators with >$20M aggregate revenue over any 12 months must sign a separate agreement; products with >100M MAU or >$20M monthly revenue must display 'Kimi K3' in the UI) license.
How much does Kimi K3 cost through the API?
$3 per 1M input tokens and $15 per 1M output tokens (USD).
Where does China AI Hub get its Kimi K3 data?
From 4 sources (official pages first), last verified 2026-09-20. Benchmark scores are labeled by source type; see the sourcing policy for details.