Loading the catalogue…
Loading the catalogue…
Moonshot AI's open-weight, 1T-parameter agentic flagship with a 262K-token context window and multimodal input, built for long-context, tool-driven Inference workflows.
Overview. Kimi K2.6 is Moonshot AI's open-weight flagship in the Kimi K2 family: a 1-trillion-parameter Mixture-of-Experts model with 32 billion active parameters per token. Built on a DeepSeek-V3-style transformer backbone with an added MoonViT vision encoder, it accepts text, image, and video input and returns text, with a toggleable Thinking mode for step-by-step reasoning. Released under a Modified-MIT license with open weights and a 262,144-token context window, it's aimed at developers who want a large-context, agentic model they can self-host or route through the Smart Router.
Key strengths. The model's clearest signal is its agentic design: Moonshot built K2.6 around multi-step tool-loop workflows, and its MoE architecture (384 experts, with 8 routed plus one shared expert active per token) keeps per-token compute manageable despite the 1T parameter count. Native INT4 quantisation in compressed-tensors format reduces the memory footprint for self-hosted Inference. Vision and video input, paired with optional Thinking-mode reasoning, extend its reach beyond pure text into multimodal, multi-step tasks.
When to use it. Reach for Kimi K2.6 when a task needs a long context window — up to 262,144 tokens — combined with multi-step, tool-calling agent behaviour, such as codebase-spanning research or long-document analysis that requires several tool calls in sequence. Turn on Thinking mode when visible chain-of-thought reasoning is more valuable than an instant single-pass answer. Its image and video input support also makes it a candidate for agent workflows that mix visual and text context.
Limitations. Independent reporting points to a knowledge cutoff around April 2025, but Moonshot has not published an official system card confirming this, so treat it as approximate. Output is text-only, with no audio or image generation. At 1T total parameters served with optional Thinking mode, expect higher latency than smaller interactive models, and no vendor-confirmed maximum output length distinct from the context window has been published.
Sovereignty note. Kimi K2.6 is created by Moonshot AI, based in China. It ships as open weights under a Modified-MIT license, so it can be self-hosted for full control over data residency; the currently listed hosted endpoint runs on a separate third-party provider, so teams with strict residency requirements should verify that endpoint's jurisdiction independently.
| Inference Provider | Sovereignty | Runs in | In / out per 1M | Status | Pin |
|---|---|---|---|---|---|
| L1 | €1.25 / 5.00 | Online |
12 use cases classified by the curator. The note says why.
262K context window allows full-length document ingestion for Q&A without chunking, a direct architectural fit for long-document analysis.
Explicit tool-calling capability is the core requirement for function-calling applications; Kimi K2.6 documents this natively.
Native tool-calling support combined with a 262K context window is well-suited to multi-tool orchestration workflows requiring large working memory across tool outputs.
Tool-calling support enables agentic code-execution loops, but no SWE-Bench or coding-specific benchmark is published to grade fit beyond baseline capability.
262K context supports grounding answers across large retrieved document sets in a single pass.
Tool-calling is a prerequisite for browser-automation agents; without curated benchmarks this is a capability-based rather than evidence-graded fit.
262K context window supports maintaining coherence across long documents, a baseline architectural advantage for long-form output tasks.
Tool-calling plus a 262K context window support multi-step plans that need to track long histories, though no reasoning/agentic benchmark scores are published to confirm planning depth.
No CLOUD Act exposure and a large 262K context window make this a reasonable sovereignty-conscious option for policy document Q&A, though no compliance-specific benchmarks are published.
No embedding, retrieval, or reranking-specific capability is documented; general tool-calling LLMs are not optimised for semantic search infrastructure.
Sourced numbers, not our own evaluations. Percentiles are against the whole Stav catalogue.
Computed from aa-intelligence-index, gpqa-diamond, aa-engineering-index, aa-coding-index, livecodebench, terminal-bench-v4, ifbench, aa-lcr.
curl https://api.stav.ai/v1/chat/completions \
-H "Authorization: Bearer $STAV_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "moonshotai/kimi-k2-6",
"reasoning_effort": "high",
"messages": [{"role": "user", "content": "Hello!"}]
}'→ X-Stav-Request-Id · X-Stav-Model · X-Stav-Provider → X-Stav-Sovereignty-Level — the rung the request was served at