Kimi K3
Kimi K3 is Moonshot AI's open-weight 2.8T multimodal MoE flagship, activating about 104B parameters per token with a 1M context window and low/high/max reasoning effort; code, 96 weight shards, and the technical report are available under the Kimi K3 License.
Data sourced primarily from official releases (GitHub, Hugging Face, papers), then benchmark leaderboards, then third-party evaluators. Learn about our data methodology
| Type | Condition | Input | Output |
|---|---|---|---|
| Text | - | ¥20.00/ 1M | ¥100.00/ 1M |
| Type | Condition | Input | Output |
|---|---|---|---|
| Text | - | $3.00/ 1M | $15.00/ 1M |
| Type | TTL | Write | Read |
|---|---|---|---|
| Text | - | - | $0.300/ 1M |
Kimi K3 currently shows benchmark results led by BrowseComp (1 / 53, score 91.20), GPQA Diamond (8 / 187, score 93.50), AA-LCR (1 / 15, score 74.70). This page also consolidates core specs, context limits, and API pricing so you can evaluate the model from benchmark results and deployment constraints together.
Want a custom combination? Open the compare tool
Kimi K3 is Moonshot AI's native multimodal agentic flagship, released on July 16, 2026. It is a 2.8T-parameter MoE model with about 104B activated parameters per token and a 1,048,576-token context window. The full model weights, inference code, and technical report became publicly available on July 27 UTC / July 28 China Standard Time through Hugging Face and GitHub.
K3 has 93 layers: one dense layer, 69 Kimi Delta Attention layers, and 24 Gated MLA layers. Stable LatentMoE routes each token to 16 of 896 experts and includes two shared experts. The hidden size is 7168, with 96 attention heads and a roughly 160K vocabulary. Its MoonViT-V2 vision encoder has 401M parameters.
Quantization-aware training starts at SFT. Released weights use MXFP4 with MXFP8 activations. The Hugging Face repository contains 96 Safetensors weight shards and roughly 1.42 TiB of weight-related large files. Moonshot recommends vLLM, SGLang, and TokenSpeed for inference.
K3 targets long-horizon software engineering, knowledge work, reasoning, and visual tasks. It accepts text, image, and video inputs and produces text. Its API supports function and tool calling, strict JSON Schema output, Partial Mode, dynamically loaded tools, automatic context caching, and OpenAI/Anthropic-compatible interfaces.
Thinking is always enabled. reasoning_effort supports low, high, and max, with max as the default. Multi-turn and tool workflows must pass the complete prior assistant message back unchanged, including reasoning_content and tool_calls.
The open-weight model card reports 45 benchmarks represented here as 50 structured rows, including five paired no-tool/tool-augmented results. All K3 results use max reasoning effort and temperature 1.0. Representative scores include GPQA Diamond 93.5, DeepSWE 67.5, Terminal-Bench 2.1 88.3, BrowseComp 91.2, MCPMark-Verified 94.5, OSWorld-Verified 84.8, and OmniDocBench 91.1.
Per 1M tokens, China pricing is ¥2 cached input, ¥20 uncached input, and ¥100 output; international pricing is $0.30, $3, and $15 respectively. API access is unlocked after at least a $1 successful top-up, and account tier determines rate limits.
Code, weights, configurations, and documentation use the Kimi K3 License. It permits use, modification, distribution, sublicensing, sale, deployment, and fine-tuning, subject to notice and legal-compliance requirements. MaaS operators with more than $20M aggregate revenue in any consecutive 12 months must obtain a separate agreement before commercial use. Commercial products exceeding 100M monthly active users or $20M monthly revenue must prominently display “Kimi K3.” The license includes exceptions for internal use, official Moonshot products, and certified inference partners.
ms:// file.Hugging Face weights and model card · GitHub code and technical report · official technical blog · API quickstart · official pricing
Yes. Moonshot released the full weights, inference code, and technical report on July 27 UTC / July 28 China Standard Time. The official Hugging Face repository contains 96 Safetensors weight shards and roughly 1.42 TiB of related large files.
K3 has 2.8T total and about 104B activated parameters, 93 layers, 69 KDA plus 24 Gated MLA attention layers, and routes each token to 16 of 896 experts with two shared experts.
The Kimi K3 License permits commercial use with conditions. Large MaaS operators must obtain a separate agreement, and very large commercial products have a prominent attribution requirement. Review the full license before use.
Thinking is always on, with low, high, and max effort levels; max is the default. The API accepts text, image, and video inputs, returns text, and supports tool calling and structured output.
Per 1M tokens, China pricing is ¥2 cached input, ¥20 uncached input, and ¥100 output; international pricing is $0.30, $3, and $15 respectively. A successful top-up of at least $1 unlocks K3 API access.
The open-weight model card reports 45 benchmarks represented as 50 structured rows. Examples include GPQA Diamond 93.5, DeepSWE 67.5, BrowseComp 91.2, MCPMark-Verified 94.5, OSWorld-Verified 84.8, and OmniDocBench 91.1.
Moonshot recommends vLLM, SGLang, or TokenSpeed. The MXFP4 release is roughly 1.42 TiB, so deployment still requires large accelerator memory, high-speed interconnects, storage, and compatible serving software.
Follow DataLearner on WeChat for AI model updates and research notes.
