Kimi K2.7-Code

Kimi K2.7-Code (API model ID kimi-k2.7-code) is Moonshot AI's open-weight, coding-focused agentic model, released 2026-06-12 as a coding-specialized descendant of Kimi K2.6 (Hugging Face model card; MarkTechPost). It is a Mixture-of-Experts model — 1 trillion total parameters, 32 billion activated per token — and reasoning is mandatory: there is no non-thinking inference path, and the API rejects any temperature/top_p pair other than its fixed defaults (Kimi API platform quickstart). Weights ship under a Modified MIT license on Hugging Face; the model is also served through the Kimi API and the Kimi Code product (Hugging Face). Moonshot positions it narrowly: its own resource page recommends Kimi K2.6 instead "for general-purpose work such as writing, analysis, and conversation," reserving K2.7-Code for coding tasks (Kimi resource page).

Key specs

Values as of release (2026-06-12), re-verified 2026-07-12 against vendor docs and models.dev.

Spec	Value	Source
Context window	262,144 tokens (256K)	Kimi API quickstart; Hugging Face
Max output	32,768 tokens — fixed default and ceiling, not configurable higher (models.dev's snapshot lists 262,144, mirroring the context-window field rather than the API's documented output cap; this page follows the vendor quickstart)	Kimi API quickstart
Modalities	Text + image + video input (via a ~400M-parameter MoonViT vision encoder); text output only	Hugging Face
Thinking	Mandatory; no non-thinking mode. temperature fixed at 1.0, top_p fixed at 0.95 — other values return an API error	Kimi API quickstart
Knowledge cutoff	Not disclosed by vendor	—
Pricing, input (cache miss)	$0.95 / MTok	Kimi resource page
Pricing, input (cache hit)	$0.19 / MTok	Kimi resource page
Pricing, output	$4.00 / MTok	Kimi resource page
License	Modified MIT (open weights)	Hugging Face

Architecture, per the Hugging Face card: 384 experts (8 selected + 1 shared per token), 61 layers (including 1 dense layer), Multi-head Latent Attention with 64 heads, and a roughly 160K-token vocabulary (Hugging Face). The model ships at native INT4 quantization — the same scheme used for Kimi K2-Thinking (Hugging Face) — and is self-hostable via vLLM, SGLang, or KTransformers, but impractical without server-class multi-GPU infrastructure (DevOps.com). A separate kimi-k2.7-code-highspeed model ID is also exposed on the API platform as a distinct, faster inference variant of the same weights, which Moonshot documents at roughly 180 tokens/s output — up to 260 tokens/s in short-context scenarios — as of 2026-07-12 (Kimi API quickstart).

Kimi Code subscription pricing (annual billing, as of 2026-06-25) lists four tiers — Moderato $15/mo, Allegretto $31/mo, Allegro $79/mo, Vivace $159/mo (Kimi resource page). Some secondary coverage cites a conflicting "$19/month starting price" for the same product; treat the vendor resource-page schedule above as authoritative and re-check platform.kimi.ai/pricing before relying on an exact tier price.

Benchmark results

Independent, third-party numbers for K2.7-Code are limited to one public leaderboard as of this writing. At the 2026-06-12 release, DevOps.com reported no independent third-party results existed for K2.7-Code on SWE-bench Verified, LiveCodeBench, or GPQA Diamond (DevOps.com). Every other headline figure from the launch is Moonshot's own proprietary benchmark suite, with no independent confirmation.

Model	% resolved	As of
GPT-5.6-Sol	96.2%	2026-07-12
Claude Fable 5	95.0%	2026-07-12
Claude Opus 4.8	88.6%	2026-07-12
Grok 4.5	86.6%	2026-07-12
GLM-5.2	82.8%	2026-07-12
GPT-5.5	82.6%	2026-07-12
Claude Sonnet 5	79.6%	2026-07-12
Gemini 3.1 Pro Preview	78.8%	2026-07-12
Kimi K2.7-Code	78.2%	2026-07-12
GPT-5.3-Codex	78.0%	2026-07-12
Kimi K2.6	76.2%	2026-07-12
GPT-5.6-Terra	75.2%	2026-07-12

SWE-bench Verified, Vals AI harness, % resolved (Vals AI — SWE-bench Verified leaderboard, as of 2026-07-12). K2.7-Code sits 9th of 12 on this independent leaderboard — 0.6 points behind Gemini 3.1 Pro Preview, 0.2 ahead of GPT-5.3-Codex, and 2.0 points above its own predecessor, Kimi K2.6 (76.2%). This is the only SWE-bench Verified number on record for the model. Given DevOps.com's release-day report that no independent SWE-bench Verified result existed as of 2026-06-12 (DevOps.com), treat this score as a later, independent addition rather than a launch-day claim.

Benchmark	K2.6	K2.7-Code	Δ vs K2.6	GPT-5.5	Claude Opus 4.8	As of
Kimi Code Bench v2	50.9	62.0	+21.8%	69.0	67.4	2026-06-12
Program Bench	48.3	53.6	+11.0%	69.1	63.8	2026-06-12
MLS Bench Lite	26.7	35.1	+31.5%	35.5	42.8	2026-06-12
Kimi Claw 24/7 Bench (agentic/tool-use)	42.9	46.9	+9.3%	52.8	50.4	2026-06-12
MCP Atlas (MCP tool orchestration)	69.4	76.0	+9.5%	79.4	81.3	2026-06-12
MCP Mark Verified (tool-invocation correctness)	72.8	81.1	+11.4%	92.9	76.4	2026-06-12

Source for the table: Hugging Face model card and Kimi resource page, cross-confirmed row-for-row by DevOps.com and CometAPI. The picture is mixed, not a clean win: K2.7-Code's raw coding-benchmark deltas (+21.8%, +31.5%) are larger than its agentic/tool-use deltas (+9.3% to +11.4%), and it trails both GPT-5.5 and Claude Opus 4.8 on five of six rows — the exception is MCP Mark Verified, where it edges out Claude Opus 4.8 (81.1 vs 76.4) (DevOps.com). As of 2026-07-12, K2.7-Code carries no entry on the SWE-bench Pro eval table (OpenAI GPT-5.6 coding eval table) or the Terminal-Bench 2.1 leaderboard (tbench.ai) that this wiki tracks.

Agentic behavior notes

Moonshot's headline claim is roughly 30% lower reasoning-token usage than Kimi K2.6 at comparable or better task completion, framed around "long-horizon" agentic coding — multi-step planning, editing, running tools, debugging (Kimi resource page; MarkTechPost). DevOps.com quotes analyst Mitch Ashley: "token efficiency is a transitory challenge in agentic coding," and "the durable opportunity is inference efficiency delivered as a governable constraint inside an AI harness, where teams operate with token budgets applied at runtime" — not a vendor's self-reported reasoning-token reduction (DevOps.com).

Two constraints matter for harness integrators specifically: tool_choice only accepts auto or none (no forced single-tool call), and the assistant's reasoning_content from prior tool-call turns must be kept in context across multi-step tool calls or the API returns an error (Kimi API quickstart). CometAPI's independent writeup separately notes "mixed results on public kernels" for some practitioners despite the strong proprietary-benchmark story, flagging that many of Moonshot's reference tasks are "in-house or specific setups" (CometAPI). No system-card-style safety or alignment disclosure for K2.7-Code specifically was found as of 2026-07-12 — no published data beyond the coding/agentic benchmark tables above.

Limitations

• No independent benchmark confirmation for the launch claims. Kimi Code Bench v2, Program Bench, MLS Bench Lite, Kimi Claw 24/7 Bench, MCP Atlas, and MCP Mark Verified are all Moonshot-proprietary; DevOps.com reported no independent SWE-bench Verified, LiveCodeBench, or GPQA Diamond result at the 2026-06-12 release (DevOps.com). Only SWE-bench Verified has since gained an independent public-leaderboard number (Vals AI, as of 2026-07-12); SWE-bench Pro and Terminal-Bench still carry none.
• Mandatory thinking mode is an operational cost, not a footnote. There is no cheap, non-reasoning path for simple edits; teams that want one must fall back to Kimi K2.6 or another model (Kimi resource page). Budget for reasoning-token overhead on every call — the effective cost delta against a non-reasoning-capable competitor can be smaller than the headline per-token price suggests.
• Tool-calling rigidity. tool_choice restricted to auto/none means harnesses that need to force a specific tool call require workaround logic rather than relying on the API (Kimi API quickstart).
• Self-hosting is impractical for most teams. Even at native INT4 quantization (Hugging Face), running the 1T-parameter weights requires server-class multi-GPU infrastructure despite the open Modified MIT license (DevOps.com).
• Pricing is more complex than the headline number. The $0.95/MTok input price only applies on a cache miss; cache-hit input is $0.19/MTok, so effective cost depends heavily on prompt-cache hit rate — a mechanic independent commentary argues undercuts the "cheaper" narrative once modeled in (Nerova.ai / Handy AI; Kimi resource page).

Related

• Kimi K2.6 — the model K2.7-Code builds on and the baseline for every vendor-reported delta above.
• Claude Opus 4.8, GPT-5.5 — the two models in Moonshot's own comparison table and on the independent SWE-bench Verified leaderboard.
• SWE-bench — the one benchmark with an independent number for this model; SWE-bench Pro and Terminal-Bench have none published as of 2026-07-12.
• Model Context Protocol — MCP Atlas and MCP Mark Verified, two of Moonshot's proprietary benchmarks, specifically measure MCP tool orchestration and invocation correctness.
• context-window — the 262,144-token context spec above.