Kimi K3 is Moonshot AI's open-weight, native multimodal agentic model — a 2.8-trillion-parameter Mixture-of-Experts model (104B activated per token) built on a new Kimi Delta Attention (KDA) + Attention Residuals (AttnRes) architecture. Moonshot AI calls it "the world's first open 3T-class model," aimed at long-horizon coding, agentic knowledge work, and reasoning, with a 1,048,576-token (1M) context window.

How does Kimi K3's architecture differ from Kimi K2?

K3 moves to a Stable LatentMoE framework that activates 16 of 896 experts per token, which Moonshot AI reports as roughly a 2.5x improvement in overall scaling efficiency over K2. The attention stack mixes 69 KDA layers with 24 Gated MLA layers (96 attention heads, 7168 hidden dimension), and the model ships with native MXFP4-weight / MXFP8-activation quantization trained in from the SFT stage rather than applied after the fact.

How does Kimi K3 compare to Claude and GPT-class models on benchmarks?

On Moonshot AI's own published table (max reasoning effort, all figures self-reported), K3 leads or is competitive on several agentic and tool-use benchmarks — e.g. BrowseComp 91.2 vs. GPT-5.6 Sol's 90.4, MCPMark-Verified 94.5 vs. 92.9, Terminal-Bench 2.1 at 88.3 close behind GPT-5.6 Sol's 88.8 — while trailing on some reasoning-heavy evals like HLE-Full (43.5 vs. Claude Fable 5's 53.3) and CritPt (23.4 vs. GPT-5.6 Sol's 32.3). It's not uniformly ahead or behind any single competitor; the pattern varies by benchmark category (reasoning, coding, agentic, vision).

Is Kimi K3 free to use commercially?

Mostly, with two revenue-triggered conditions in the custom Kimi K3 License (same structure as K2's license). If you run K3 as a "Model as a Service" (giving third parties API-level access to inference/fine-tuning) and your combined revenue exceeds $20M USD over any 12 consecutive months, you need a separate commercial agreement with Moonshot AI. Separately, if K3 powers a product with over 100M monthly active users or over $20M/month in revenue, you must prominently display "Kimi K3" in that product's UI. Internal use and end-user products that merely embed the model's capabilities are exempt from both.

What inference engines support Kimi K3?

Moonshot AI recommends vLLM (with published recipes), SGLang (with a cookbook), and TokenSpeed. There's also a hosted API at platform.kimi.ai that's OpenAI/Anthropic-compatible, selectable as the `kimi-k3` model. For an agent harness, Moonshot AI points to their own Kimi Code CLI, switchable via the `/model` command.

Does Kimi K3 support extended "thinking" like reasoning models?

Yes — thinking is always on, returned as a `reasoning_content` field, with effort configurable via a `reasoning_effort` parameter ("low", "high", or "max", defaulting to "max"). For multi-turn conversations, the API requires the full prior assistant message — including `reasoning_content` and any `tool_calls` — to be passed back as-is on the next turn, not just the final `content` string.

Kimi K3: Moonshot AI's 2.8T-Parameter Open-Weight Frontier Model

1M Context Window LLM 2026: Real Tests • Open Interpreter: A Codex Fork That Emulates Claude Code and Kimi’s Harness

Kimi K3 — from github.com/MoonshotAI/Kimi-K3

What Is Kimi K3? #

Kimi K3 is Moonshot AI’s newest open-weight model, and per the README, their most capable to date: a 2.8-trillion-parameter Mixture-of-Experts model with 104B activated parameters per token, built on a new Kimi Delta Attention (KDA) and Attention Residuals (AttnRes) architecture. Moonshot AI describes it as “the world’s first open 3T-class model” — native multimodal (text, image, video-frame understanding), a 1,048,576-token context window, and released with full weights under a custom Kimi K3 License.

🔗 GitHub: https://github.com/MoonshotAI/Kimi-K3 🤗 Weights: huggingface.co/moonshotai/Kimi-K3 📄 Tech report: linked from the repo as k3_tech_report.pdf

At 8,100+ GitHub stars, created July 27, 2026 and pushed as recently as August 6, 2026.

Architecture: What Changed From K2 #

Spec	Value
Total parameters	2.8T
Activated parameters	104B
Layers	93 (1 dense + 92 MoE)
Attention composition	69 KDA + 24 Gated MLA
Attention heads	96 (hidden dim 7168)
Experts	896 total, 16 selected/token, 2 shared
Vocabulary	160K tokens
Context length	1,048,576 tokens
Vision encoder	MoonViT-V2 (401M params)
Quantization	Native MXFP4 weights / MXFP8 activations (quantization-aware trained)

Per Moonshot AI, the Stable LatentMoE framework activating 16 of 896 experts yields “an approximate 2.5× improvement in overall scaling efficiency” over K2. The quantization detail is worth flagging separately: MXFP4/MXFP8 is trained in from the SFT stage onward, not bolted on as a post-hoc compression step — the stated goal being broad hardware compatibility without a separate quantization pass degrading quality.

Benchmark Highlights (Self-Reported, Max Effort) #

Moonshot AI’s README publishes a large comparison table against Claude Fable 5, Claude Opus 4.8, GPT-5.6 Sol, GPT-5.5, and GLM-5.2. A representative slice — K3 doesn’t sweep every category, and the pattern shifts by benchmark type:

Benchmark	Kimi K3	Best of the rest
BrowseComp (agentic web)	91.2	GPT-5.6 Sol 90.4
MCPMark-Verified (MCP tool use)	94.5	GPT-5.6 Sol / GPT-5.5 tied 92.9
Terminal-Bench 2.1	88.3	GPT-5.6 Sol 88.8
GPQA Diamond (reasoning)	93.5	GPT-5.6 Sol 94.1
HLE-Full	43.5	Claude Fable 5 53.3
CritPt (physics reasoning)	23.4	GPT-5.6 Sol 32.3
Video-MME (w/ subtitles)	90.0	GPT-5.6 Sol 89.5
OmniDocBench (document vision)	91.1	Claude Fable 5 89.8
Harvey Lab-AA (legal)	94.6	Claude Fable 5 93.6

Read this carefully: these are Moonshot AI’s own numbers from their tech report, not third-party reproductions. K3 leads on agentic tool-use and several vision/document benchmarks, but trails Claude Fable 5 and GPT-5.6 Sol on some of the hardest pure-reasoning evals (HLE-Full, CritPt). Treat it as “strong, benchmark-dependent,” not “best at everything.”

Deployment and Model Usage #

Recommended inference engines, per the README:

vLLM — published recipes at recipes.vllm.ai
SGLang — cookbook at docs.sglang.io
TokenSpeed — recipes at lightseek.org

A hosted, OpenAI/Anthropic-compatible API is available at platform.kimi.ai (model name kimi-k3).

Thinking is always on. Reasoning effort is set via a reasoning_effort field ("low" / "high" / "max", default "max"), returned as a separate reasoning_content field. The one gotcha worth flagging for anyone integrating this: K3 was trained in preserved thinking history mode, meaning multi-turn calls must pass the complete prior assistant message back — reasoning_content and tool_calls included, not just the final content string — or the model loses the thread on follow-up turns.

For an agent harness, Moonshot AI points to their own Kimi Code CLI — run it in a terminal and switch to K3 with the /model command.

Licensing: What Triggers a Paid Agreement #

The Kimi K3 License (custom, same shape as K2’s) is permissive by default — free to use, modify, fine-tune, and redistribute — with two revenue-gated conditions:

“Model as a Service” threshold: if you (plus affiliates) give third parties API-level control over K3’s inputs/parameters/fine-tuning and your combined revenue exceeds $20M USD over any 12 consecutive months, you need a separate commercial agreement with Moonshot AI.
Attribution at scale: if K3 powers a product with 100M+ monthly active users or $20M+/month revenue, “Kimi K3” must be prominently displayed in that product’s UI.

Both conditions are waived for internal use and for end-user products that merely embed K3’s capabilities without exposing model-level control to third parties.

License #

Kimi K3 License (custom, permissive with revenue-based commercial terms) — see LICENSE.