Kimi K2.6

Kimi K2.6 (model ID kimi-k2.6) is Moonshot AI's flagship open-weight model: a 1-trillion-parameter Mixture-of-Experts (MoE) architecture that activates 32 billion parameters per token (Hugging Face model card, accessed 2026-07-12). Moonshot shipped it as a general-availability release, removing the earlier "Code Preview" label across kimi.com, the Kimi App, the official API, and the Kimi Code CLI (official model page, accessed 2026-07-12). Moonshot's own forum announcement is dated 2026-04-21 (Moonshot AI official forum, accessed 2026-07-12); independent tech press reported the release a day earlier, on 2026-04-20 (MarkTechPost, accessed 2026-07-12). This page uses the vendor's own announcement date, 2026-04-21, which also matches the models.dev registry snapshot (accessed 2026-07-12). K2.6 keeps the MoE skeleton of its predecessor, Kimi K2.5; the delta is concentrated in post-training — more compute on long-horizon stability, instruction following, and swarm coordination, plus native INT4 quantization-aware training (kimi.com blog, accessed 2026-07-12). Moonshot later shipped a coding-specialized sibling, Kimi K2.7 Code; this page covers K2.6 only.

Key specs

Values as of the 2026-04-21 release, re-verified 2026-07-12 against vendor docs and the models.dev registry snapshot.

Spec	Value	Source
Release date	2026-04-21 (official announcement; independent press dated the same release 2026-04-20)	official forum; models.dev
Total / activated parameters	1 trillion total (MoE), 32 billion activated per token	Hugging Face model card
Architecture	61 layers (1 dense), Multi-head Latent Attention, 384 experts (8 routed + 1 shared per token), vocab 160,000, SwiGLU, ships with native INT4 QAT	Hugging Face model card
Context window	262,144 tokens (~256K), uniform across all K2.6 variants	Hugging Face model card; models.dev
Max output	32,768 tokens (API default)	quickstart docs
Modalities	Text, image, video input via a native 400M-parameter MoonViT vision encoder (no separate vision module); text output	Hugging Face model card
Knowledge cutoff	Not disclosed in any primary Moonshot source reviewed for this page	—
License	Modified MIT — UI attribution ("Kimi K2") required only above 100M monthly active users or $20M/month product revenue	HF LICENSE
Input price (cache miss)	$0.95 / Mtok	pricing docs
Input price (cache hit)	$0.16 / Mtok (~83% off)	pricing docs
Output price	$4.00 / Mtok	pricing docs

The models.dev registry snapshot lists max output equal to the full context window (262,144 tokens); this page follows the vendor's own API guide instead, which states a 32,768-token default — treat the higher registry figure as unconfirmed until Moonshot's docs say otherwise. The context window itself is uniform whether K2.6 is called through kimi.com, the API, or a self-hosted deployment via vLLM, SGLang, or Moonshot's own KTransformers inference engine (Hugging Face model card, as of 2026-07-12). Prompt caching cuts input cost by roughly 83% on a cache hit — for agent-team designs that re-send large system prompts or tool schemas on every turn, cache-hit rate is a bigger cost lever than the base token price (pricing docs, as of 2026-07-12). The license's UI-attribution clause only activates at product scale (100M+ MAU or $20M+/month revenue) — irrelevant for almost all individual or team usage (HF LICENSE, accessed 2026-07-12).

Benchmark results

Benchmark	K2.6 score	Comparison	Source	As of
SWE-bench Verified (Vals AI harness, independent)	76.2%	Claude Fable 5 95.0%, Claude Opus 4.8 88.6%, Kimi K2.7 Code 78.2% — same leaderboard	Vals AI	2026-07-12
SWE-bench Verified (Moonshot self-reported)	80.2%	—	HF model card	2026-04 release, accessed 2026-07-12
SWE-bench Pro (Moonshot self-reported)	58.6%	GPT-5.4 57.7%, Claude Opus 4.6 (max effort) 53.4%, Gemini 3.1 Pro 54.2%, Kimi K2.5 50.7% — according to Moonshot	official forum	2026-04 release, accessed 2026-07-12
Terminal-Bench 2.0 (Terminus-2 harness, Moonshot self-reported)	66.7%	Claude 4.6 65.4%, Kimi K2.5 50.8% — according to Moonshot	kimi.com blog	2026-04 release, accessed 2026-07-12
LiveCodeBench v6 (Moonshot self-reported)	89.6%	—	HF model card	2026-04 release, accessed 2026-07-12
Humanity's Last Exam, full set with tools (Moonshot self-reported)	54.0%	GPT-5.4 52.1%, Claude Opus 4.6 53.0%, Gemini 3.1 Pro 51.4% — according to Moonshot	official forum	2026-04 release, accessed 2026-07-12
BrowseComp (Moonshot self-reported)	83.2%	Gemini 3.1 Pro 85.9% — according to Moonshot	kimi.com blog	2026-04 release, accessed 2026-07-12

Full leaderboard: Vals AI — SWE-bench Verified leaderboard, as of 2026-07-12.

Reading notes for practitioners:

• The independently reproduced score sits below the vendor-reported one. Vals AI's own harness resolves 76.2% of SWE-bench Verified with K2.6, about 4 points under Moonshot's self-reported 80.2% (Vals AI; HF model card, both accessed 2026-07-12) — a harness gap most vendor release blogs don't surface. Treat the Vals AI number as the more load-bearing one for a buy/adopt decision.
• SWE-bench Verified and SWE-bench Pro are distinct benchmarks; their scores are not comparable to each other, and neither is comparable to the Terminal-Bench row below.
• The Terminal-Bench row uses version 2.0, not the 2.1 methodology on the current tbench.ai leaderboard (accessed 2026-07-12) — that leaderboard carries no K2.6 entry as of 2026-07-12, so the 66.7% figure has no independent cross-check available on this page.
• Moonshot's release materials describe broad improvement over K2.5 across every compared benchmark — K2.6 leads K2.5 on each row of the release table, including both Humanity's Last Exam variants — and report a +185% gain on long-horizon multi-file coding tasks — no breaking API changes accompany the upgrade, so it is a drop-in model-name swap from K2.5 (kimi.com blog, accessed 2026-07-12).

Agentic behavior notes

• Agent Swarm is the headline feature: K2.6 scales to 300 domain-specialized parallel sub-agents executing up to 4,000 coordinated steps in a single autonomous run, up from K2.5's 100 sub-agents and 1,500 steps (kimi.com blog; official forum, both accessed 2026-07-12). The differentiator is parallel/orchestrated throughput, not single-agent reasoning depth — the benchmark table above shows K2.6 is competitive but not clearly ahead of top-tier closed models on single-agent coding tasks.
• Long-horizon execution. Moonshot documents 4,000+ tool calls sustained over 12+ continuous hours with reported generalization across Rust, Go, and Python, and cites a 5-day autonomous run inside third-party "proactive agent" frameworks (OpenClaw, Hermes) (kimi.com blog, accessed 2026-07-12).
• Drop-in harness compatibility. The API is an OpenAI-and-Anthropic-compatible chat-completions interface (streaming, tools/function-calling, tool_choice, temperature, top_p, max_tokens), which is how K2.6 plugs into Claude Code, OpenCode, Cline, Roo Code, OpenClaw, and Hermes Agent as an alternate backend model without a harness rewrite (quickstart docs, accessed 2026-07-12).
• Thinking-mode tool-calling constraint. When thinking is enabled, tool_choice is restricted to auto/none (no forcing a specific tool call), reasoning_content must be retained across multi-step tool calls, and the web-search tool is incompatible with thinking mode — a concrete integration detail for anyone wiring K2.6 into a custom harness rather than an off-the-shelf one that already handles it (quickstart docs, accessed 2026-07-12).

Limitations

• No independent reproduction for most reported scores. Outside SWE-bench Verified (where Vals AI's independent number trails Moonshot's own by about 4 points, see Benchmark results), every other benchmark figure on this page is Moonshot's self-measurement, including its comparison numbers for competitor models (kimi.com blog, accessed 2026-07-12).
• Knowledge cutoff undisclosed. No primary Moonshot source reviewed for this page states a training-data cutoff date; treat any specific date circulating elsewhere as unverified.
• Independent safety evaluation exists only for the predecessor. An independent academic safety evaluation (published 2026-04-03) covers Kimi K2.5, not K2.6: dual-use capability on CBRNE-adjacent prompts comparable to GPT-5.2/Claude Opus 4.5 with fewer refusals, competitive but not frontier-level autonomous cyber-offense capability, concerning sabotage-ability and self-replication propensity in agentic settings without evidence of persistent malicious goals, and higher compliance with disinformation/copyright-infringing requests than some peers. Moonshot has not published an equivalent evaluation for K2.6 itself, so these findings should be read as predecessor-model context, not a K2.6 result.
• Launch communication was rocky. Practitioner reaction on Hacker News to the initial K2.6 preview centered on confusing, contradictory pricing-tier documentation and the absence of an initial dedicated blog post — a minor vendor-communication data point, not a capability finding (Hacker News thread, accessed 2026-07-12).

Related

• Kimi K2.7 Code — coding-specialized sibling Moonshot shipped after K2.6; check it directly before assuming K2.6 is Moonshot's current best pure-coding option.
• Claude Code, agent harness — integration targets for K2.6 via its OpenAI/Anthropic-compatible API.
• Sub-agent — the unit Agent Swarm parallelizes to 300 per run.
• Claude Opus 4.8 — the closed-model baseline in the independently reproduced SWE-bench Verified row above (Vals AI harness); Moonshot's own self-reported comparison rows elsewhere on this page benchmark against Claude Opus 4.6 and GPT-5.4, not against Opus 4.8 or GPT-5.5.
• Evaluation awareness, benchmark contamination — relevant context for reading any vendor's self-reported comparison numbers, including Moonshot's here.