更新
What is Kimi K3?
Kimi K3 is Moonshot AI's flagship language model, released on July 16, 2026, built for long coding runs, agentic web research, and document work. It is a from-scratch architecture rather than a K2 revision: 2.8 trillion parameters with 16 of 896 experts active per pass, and a hybrid linear attention design called Kimi Delta Attention paired with Attention Residuals. That attention work is the reason the 1M-token window is usable rather than decorative. On UnificAlly it runs as moonshotai/kimi-k3 with a 1,048,576 token context, 131,072 output tokens by default, and thinking on every request.
Key features of Kimi K3
#14 on Arena English Text
K3 ranks #14 of 414 models on Arena's English Text board at 1494±6, level with Claude Sonnet 5.5 at 1495 and Claude Opus 5 at 1494, as of October 13, 2026. The rating is no longer preliminary at 11,567 votes. The only open-weights model above it is MiMo V2.6 Pro at #12, on just 1,505 votes.
A 47-place jump in one generation
Kimi K2.6 sits at #61 with 1469 Elo. K3 sits at #14 with 1494, 25 points and 47 places higher, while the context window goes from 262k to 1M tokens.
1M token context, 131k default output
The window holds a large repo, a long filing pack, or a full agent transcript at once. A single response defaults to 131,072 tokens and can be raised to the full 1M when you need one long artifact instead of many calls.
Full agent and tool surface
K3 exposes the whole tool-calling surface through the same endpoint: tool definitions, tool_choice including required, dynamic tool loading for agents carrying a large tool set, strict json_schema structured output, and partial mode. Streaming returns reasoning and answer text as separate deltas, so you can show thinking without parsing it out of the response.
Fourth on the Artificial Analysis index
Independent testing puts K3 at #4 of 187 models on the Intelligence Index at a score of 57. The DeepSWE leaderboard, which runs every model on one shared harness, adds a second outside check: 69% for K3, against the 67.5 Moonshot reported with its own harness.
Open weights since July 27, 2026
Moonshot published the full weights on Hugging Face on July 27, as promised, with a technical report. They ship natively quantized, MXFP4 weights with MXFP8 activations, under the Kimi K3 License, which allows commercial use and fine-tuning. K3 is the first open model in the 3-trillion class, though 2.8T parameters means a multi-node deployment, not a workstation.
What Kimi K3 is best for
Long coding sessions
Multi-hour work across a repo where the model has to hold context and keep going without supervision.
Agentic web research
Multi-step search and synthesis where the model plans queries, reads results, and follows leads.
Document and spreadsheet work
Long filings, contracts, and structured files read in one pass and returned as a brief.
Tool-heavy agents
Strict tool choice and dynamic tool loading, for agents that carry a large tool surface.
Single-artifact generation
Reports, migrations, and drafts that need one very long response instead of stitched calls.
Kimi K3 use cases
Build a coding agent that plans a migration, edits files across a repo, runs tests, and keeps going across a long session without restarting. Use it as the research layer in a product that has to plan queries, read what comes back, and synthesise an answer with sources. It also fits document intelligence work where long contracts, filings, and spreadsheets are read together and returned as structured output. For teams already running an OpenAI-compatible agent loop, K3 slots in as a model id change.
Kimi K3 limitations
Thinking is always on and reasoning_effort defaults to max. Moonshot added low and high after launch, but a request that leaves the field out pays full reasoning cost, so simple chat and extraction carry latency and token spend they do not need unless you ask for less.
Speed is a real constraint. Artificial Analysis measures 62 output tokens per second, below the median for models in its class.
The model is sensitive to thinking history. Moonshot trained it in a preserved-history mode, and a harness that fails to pass earlier reasoning content back, or a session switched to K3 from another model mid-run, can produce unstable output.
It also acts rather than asks. On ambiguous instructions K3 makes unrequested decisions, so constraints belong in the system prompt rather than in follow-up corrections.
Moonshot says plainly that overall performance still trails Claude Fable 5 and GPT 5.6 Sol, and most of the benchmark table it published uses its own in-house harness while competitor rows use other vendors' tooling. Treat those self-reported numbers as unconfirmed.
Kimi K3 vs Claude Fable 5
On Arena's English Text board the two are closer than anywhere else: as of October 13, 2026, K3 sits at #14 with 1494±6 against Fable 5 at #3 with 1513, a 19-point gap. The separation shows up elsewhere. Fable 5 runs adaptive thinking, so it spends reasoning only where a task needs it, while K3 runs at max unless you pick a lower effort per call. Fable 5 also takes image input alongside text through the same endpoint. Moonshot itself puts Fable 5 ahead on overall user experience. Pick K3 for long autonomous coding and web research runs where token cost per unit of work matters, or when you want weights you can run yourself; pick Fable 5 when you want the stronger all-round model and want it to judge how much to think.
When to use Kimi K3
Use Kimi K3 for work that runs long and mostly unattended: repo-scale coding, multi-step research, and document passes that would otherwise be split across many prompts. It is a poor fit for chat, short extraction, and anything latency-sensitive, because it thinks at full effort by default and generates slower than most models in its class.
Kimi K3 API examples
Call Kimi K3 by POSTing to /v1/chat/completions, the OpenAI-compatible endpoint shared by every language model in the UnificAlly API. Parameter docs live at docs.unifically.com/models/llm/moonshotai/kimi-k3.
curl -X POST https://api.unifically.com/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "moonshotai/kimi-k3",
"messages": [
{ "role": "user", "content": "Plan a migration from Express to Fastify for this service and list the risks in order." }
]
}'
The response comes back synchronously with the completion. Set "stream": true to receive tokens as they generate; thinking and answer text arrive as separate deltas.
FAQ
Kimi K3に関するよくある質問
moonshotai/kimi-k3, called through the OpenAI-compatible POST /v1/chat/completions endpoint with your UnificAlly API key.
On UnificAlly, $2.25 per 1M input tokens, $11.25 per 1M output tokens, and $0.225 per 1M cached input tokens. Context caching is automatic, and reasoning tokens bill as output. Live rates are on the pricing page.
A 1,048,576 token context window. A single response defaults to 131,072 output tokens and can be raised to the full 1M.
No, thinking runs on every request, but you can turn it down. Moonshot documents reasoning_effort values low, high, and max for K3, and max is the default when the field is left out, so a call that sets nothing pays for full reasoning.
Yes. Moonshot published the full weights on Hugging Face as moonshotai/Kimi-K3 on July 27, 2026, with a technical report, under the Kimi K3 License. Commercial use and fine-tuning are allowed, but a model-as-a-service business with more than 20 million US dollars in revenue over 12 months needs a separate agreement with Moonshot.
Yes. Tool definitions, tool_choice including required, dynamic tool loading, and strict json_schema structured output all work through the standard chat completions request. Partial mode is available for prefilling an assistant turn.
2.8 trillion parameters, with 16 of 896 experts active per forward pass under Moonshot's Stable LatentMoE framework. Sparsity is what keeps a model that size servable.
No. Artificial Analysis measures 62 tokens per second, below the median for its class, and at the default max effort every request pays full reasoning cost. Pick it for hard work, not for chat latency.
As of October 13, 2026, Kimi K3 is #14 of 414 models on Arena's English Text board at 1494 Elo on 11,567 votes, 19 points behind Claude Fable 5 at #3. Its rating is no longer marked preliminary.
Two things. K3 is sensitive to thinking history, so a harness that drops earlier reasoning content can destabilise output quality mid-session. It also acts on ambiguous instructions rather than asking, so constrain it explicitly in the system prompt.
Kimi K3のガイドと比較
このモデルを実行する前に、料金、制限、テスト済みの出力をご確認ください。
