Kimi K3

Kimi K3 is a sparse mixture-of-experts model with 2.78T total parameters and 104.2B activated per token. It has 93 layers (the first dense), a hidden size of 7,168, 96 attention heads, a latent MoE dimension of 3,584, a per-expert hidden size of 3,072, and 896 routed experts, of which 16 are selected per token alongside 2 shared experts. Attention is hybrid: 69 layers use Kimi Delta Attention, a gated linear-attention mechanism whose state does not grow with sequence length, and every fourth layer plus the final one (24 in total) uses gated Multi-head Latent Attention. The model has no explicit positional embedding. KDA's gating and decay encode position, which is why it extends to 1M tokens without RoPE rescaling. Pre-training ran at 8K then 64K context, and the cooldown extended it from 256K to 1M. The vision encoder, MoonViT-V2 (401M parameters, 27 layers), is trained jointly with the language model from the start rather than attached afterwards, and one multi-token-prediction layer doubles as a speculative-decoding draft. The vocabulary is 160K tokens (163,840 in the config) using a tiktoken tokenizer. Weights are MXFP4 from quantization-aware training and total about 1.56 TB across 96 shards. Training-token count, training compute and knowledge cutoff are undisclosed. Moonshot recommends serving on supernodes of 64 or more accelerators and has contributed a KDA prefix-caching implementation to vLLM. That caching is how it sells the full 1M window at one flat price.

Coding (9.3) rests on a mix of vendor and independent results. Moonshot reports 88.3 on Terminal-Bench 2.1 (GPT-5.6 Sol 88.8, Fable 5 88.0), 67.5 on DeepSWE v1.1 (67.3 on the official leaderboard's shared mini-SWE-agent harness) and 42.0 on SWE-Marathon. Independently, K3 is the top open-weights model in LMArena's Text Coding category (1538, rank 7, LMArena, 2026-09-13) and on its separate WebDev arena (1674). The discount from a higher score comes from Vals AI's ProgramBench run: a 62.77% raw pass rate against Claude Opus 5's 82.3% and Sol's 77.6%, with 22 of 200 tasks scoring zero, mostly because the code did not build. Z.ai's GLM-5.3 card also reports K3 at 17.4 on Terminal-Bench 3.0, against 28.3 for GLM-5.3 and 34.6 for Sol. Reasoning (9.3) is frontier-level on Moonshot's numbers: GPQA Diamond 93.5 and HLE 43.5 without tools (56.0 with tools). Math (9.0) has no published AIME or MATH result. The score leans on MathVision 94.3 and CritPt 23.4, which trails Sol's 32.3. Agentic work (9.3) is the other headline: BrowseComp 91.2, MCP-Atlas 84.2, Toolathlon-Verified 76.5 and OSWorld-Verified 84.8, all vendor-run. Moonshot itself warns that K3 can be "excessively proactive" and make unexpected decisions on the user's behalf. Long context (9.0) is native rather than extended: 74.7 on AA-LCR (as cited by Moonshot) and 90.4 on BrowseComp using the full 1M window with no context management. Vision (9.0) and document work (9.0) are genuine strengths for an open model: MMMU-Pro 81.6, Video-MME 90.0, and OmniDocBench 91.1, ahead of Fable 5's 89.8 in the same table. Multilingual (8.5) is strong in English and Chinese, with no multilingual benchmark published. Creative writing (8.0) has no direct benchmark, but LMArena Text places it level with GPT-5.6 Sol. Instruction following (8.3) is held back by that proactiveness and by instability when the thinking history is not passed back. Function calling (9.0) is well supported: MCPMark-Verified 94.5, tool_choice: "required", dynamic tool loading and strict JSON Schema output. Safety calibration (6.0) reflects the UK AISI / CAISI finding that K3's safeguards did not stop exploit development, together with the absence of a published safety framework. Real-time data (4) is limited: a server-side web_search tool exists in the Responses API, but Moonshot's K3 guide says web search is being updated and is not recommended for production for now.

Every row sourced to the model card is vendor-reported. Moonshot ran K3 at maximum reasoning effort, often in its own Kimi Code harness, while other models ran in Claude Code or Codex, and several competitor figures come from third-party leaderboards. The K2.7-Code figures come from that model's own card. Moonshot cites GDPval-AA, AA-LCR, CritPt and SciCode from Artificial Analysis as of 2026-07-23. Its technical report also relays an Artificial Analysis Intelligence Index v4.1 score of 57.1 (fourth of 580 models at the time) and a Vals Index of 74.7%. Both are vendor relays and are not entered in this page's comparable columns. The Vals AI ProgramBench row and the three LMArena rows are independent (LMArena Text and its Coding category as of 2026-09-13, with 20,987 and 5,507 votes; the separate WebDev arena as of 2026-09-11, with 4,547 votes). On all three K3 is the highest-rated model under an open-weights license. The catalog columns carry GPQA Diamond, HLE (no tools), Terminal-Bench (2.1) and MMMU (MMMU-Pro, no tools) from Moonshot's card, and LMArena's Text board and its Coding category for the two Elo columns; the WebDev arena uses a different scale and stays out of those columns.

Speed is the model's weakest practical dimension, and no independent throughput measurement is cited on this page, so the speed fields stay empty. The only published rate is secondhand. Z.ai's GLM-5.3 model card rescales one of its benchmarks using Artificial Analysis per-model rates of 40 tokens/sec for Kimi K3, 47 for Qwen3.8-Max and 115 for GLM-5.3. That is consistent with the architecture: 104B active parameters per token is about three times Kimi K2's 32.6B, and the API defaults to maximum reasoning effort. Set reasoning_effort: "low" for interactive use, and choose the level before a session starts, because switching levels invalidates prefix-cache hits. Moonshot serves K3 on its Mooncake disaggregated-inference stack and reports cache-hit rates above 90% on coding workloads, which cuts cost more than latency. For raw speed, Moonshot's own fast option is still kimi-k2.7-code-highspeed (about 180 tokens/sec per Moonshot).

At list price K3 sits between the open flagships and the closed frontier. Per 1M tokens it costs 2.1x GLM-5.3 on input and 3.4x on output ($1.40/$4.40), 1.5x and 2.5x Qwen3.8-Max ($2.00/$6.00, Singapore), and 2.3x and 3.8x DeepSeek V4-Pro at peak rates ($1.32/$3.96). It is 25% cheaper than GPT-5.6 Sol's promotional $4.00/$20.00 and 40% cheaper than Claude Opus 5's $5.00/$25.00. Two things raise the effective cost: the default max reasoning effort produces long outputs, and there is no batch discount. Caching is the counterweight, because agent loops with a stable prefix pay $0.30 for most of their input. The Kimi Open Platform deducts usage 50% from cash and 50% from vouchers.

First-party API. The global platform (platform.kimi.ai, https://api.moonshot.ai/v1) exposes K3 through OpenAI-compatible Chat Completions and Responses APIs and an Anthropic-compatible Messages API, so Claude Code, Codex and OpenCode can point at it directly. A separate China platform uses api.moonshot.cn.

Hosted by others. Hugging Face lists Together AI, Fireworks AI, Baseten and Featherless AI as live inference providers for the model; DeepInfra shows an error status there. No managed listing on Bedrock, Vertex AI or Azure is documented by Moonshot.

Self-hosting. The weights (moonshotai/Kimi-K3) are about 1.56 TB in MXFP4, which is more than a single eight-GPU node of 80 GB or 141 GB cards can hold. Moonshot recommends 64-plus-accelerator supernodes and supports vLLM, SGLang and TokenSpeed. Moonshot does not state a minimum VRAM figure.

License: Kimi K3 License. It is MIT-derived: it grants the rights to use, copy, modify, merge, publish, distribute, sublicense and sell the weights, code and documentation, and to run, deploy, fine-tune and create derivative works, subject to four sections:

1. Keep the copyright and permission notice, and comply with applicable law.
2. If the licensee or any affiliate operates a Model-as-a-Service business, and the combined revenue of the licensee and its affiliates exceeds US$20 million over any consecutive 12 months, the licensee must sign a separate agreement with Moonshot AI before using K3 or a derivative for any commercial purpose. The revenue test counts total revenue, not just MaaS revenue. Model-as-a-Service means giving a third party access to inference or fine-tuning with meaningful control over inputs, parameters or training data. It excludes end-user products that embed the model only inside specific features or harnesses, and it excludes simply relaying requests to models hosted by others.
3. Commercial products or services with more than 100 million monthly active users or more than US$20 million in monthly revenue must prominently display "Kimi K3" in their user interface.
4. Sections 2 and 3 do not apply to internal use, meaning nothing (the model, its outputs or its capabilities) is made available to third parties. They also do not apply to use through Moonshot's official products or its "certified inference partners", which the license does not list.

The warranty disclaimer is the standard as-is one. The MaaS threshold is the strictest among the open 2T-class flagships: Qwen3.8's comparable clause starts at US$50 million, and GLM-5.3's at US$10 billion (a security review rather than a separate agreement).

Moonshot publishes no safety framework, risk-tier label or compliance certification for K3. The Kimi Open Platform terms (updated 2026-07-30) let Moonshot use customer content, meaning inputs and outputs, to provide, develop and improve its services. Customers who want to restrict its use for training or improving Moonshot models have to arrange enterprise terms or a separate written agreement, and without one the content may be used. Moonshot's Zero Data Retention page states that enterprise customer data is not used for training by default. ZDR itself is available on request to enterprise customers: prompts and responses are deleted once the response is returned, but images and videos sent as direct file uploads are not covered. The global platform is operated by Moonshot AI Pte. Ltd. from servers in Singapore, under Singapore law with SIAC arbitration. Automated content-safety reviews apply to all API traffic, including under ZDR. The UK AI Security Institute and the US Center for AI Standards and Innovation published a joint preliminary assessment on 2026-07-23. It found that K3's safeguards did not prevent attempts at exploit development or offensive cyber operations. It also found that K3's cyber capability sits well below the leading US closed models: 32% on ExploitBench (GLM-5.2: 24%), arbitrary code execution on 0 of 41 tasks where the most cyber-capable models average 20, and step 17 of a 32-step simulated corporate network attack against 28.5 for the top US models, with one full solve in ten attempts. Moonshot's own report reaches consistent conclusions and describes 16 previously unknown vulnerabilities K3 found in real projects. Anyone self-hosting takes on every safeguard themselves.

K3 is the default model across Moonshot's own products: the Kimi app (iOS, Android, HarmonyOS, web), the Kimi Work desktop app (Widgets and Dashboard), and Kimi Code, the terminal agent Moonshot recommends as K3's harness. Moonshot publishes setup guides for Claude Code, Codex, OpenCode and Hermes Agent, and the OpenAI and Anthropic SDKs work against its endpoints in Python and TypeScript. For self-hosting, vLLM, SGLang and TokenSpeed have K3 recipes, and Moonshot contributed KDA prefix caching to vLLM. Hugging Face routes K3 through Together AI, Fireworks AI, Baseten and Featherless AI. Adoption is fast for a model this large: about 1.9 million downloads and 11,000 likes on Hugging Face, and nearly 21,000 LMArena Text votes. Enterprise uptake outside China is still early.