Kimi coding model for Coding / Agent work
Officially positioned as Kimi's strongest coding model so far; K2.7 Code cannot disable thinking.
Overview
Kimi K2.7 Code is the model Kimi's developer docs now emphasize for Coding / Agent work. The official guide describes it as Kimi's strongest coding model so far: it follows instructions more reliably in long contexts, completes programming tasks with higher success rates, accepts text, image, and video input, and returns text for both chat and agent workflows.
Compared with Kimi K2.6, K2.7 Code concentrates the upgrade on long-horizon coding, autonomous agent execution, and interactive programming feel. Kimi also offers kimi-k2.7-code-highspeed: the same model served in a faster tier, with output speed roughly 5-6x the standard version, around 180 Tokens/s in typical coding contexts and up to 260 Tokens/s in short-context cases. CrossModel serves the standard API model id: kimi-k2.7-code.
Key capabilities
| Dimension | Detail |
|---|---|
| Context window | 262,144 tokens (~256K) |
| Default generation cap | 32,768 tokens (max_tokens default) |
| Input modalities | Text, image, video |
| Output modalities | Text |
| Tools | streaming, JSON Mode, Partial Mode, ToolCalls, automatic context caching |
K2.7 Code defaults to thinking and does not allow thinking to be disabled; requests that turn it off fail.
temperatureis fixed at 1.0 andtop_pat 0.95. With tool use,tool_choicemust beautoornone, and multi-step tool calls must preserve the previous assistant message'sreasoning_contentin context. Current pricing is in the model catalog.
Benchmarks
Kimi concentrates K2.7 Code's gains on coding and long-horizon engineering, not single-turn answer quality. The official benchmark chart also shows K2.6, GPT-5.5, and Opus 4.8 references; the direct K2.6 deltas below are the cleanest way to read the upgrade.
A clear lead over K2.6 on the coding axis
All comparisons are official K2.7 Code vs K2.6 numbers.
Against K2.6, K2.7 Code rises from 50.9 to 62.0 (+21.8%) on Kimi Code Bench v2, from 48.3 to 53.6 (+11.0%) on Program Bench, and from 26.7 to 35.1 (+31.5%) on MLS Bench Lite. All three are programming-focused measurements, so the advantage maps most directly to writing code, editing code, and sustaining engineering work over long contexts.
Agentic capability and HighSpeed
Steady agent gains, lower wait time on the fast tier
Agent comparisons come from the official benchmark chart; HighSpeed is the same model as the standard tier.
On the agentic axis, K2.7 Code also improves steadily over K2.6: Kimi Claw 24/7 Bench 42.9 -> 46.9, MCP Atlas 69.4 -> 76.0, and MCP Mark Verified 72.8 -> 81.1, putting the gains in roughly the 9% to 11% range. For coding agents, these measurements matter more than single-shot code completion because they test stability across tools, context, and multi-step execution.
The HighSpeed version addresses a different need: lower waiting time when cost is less important than interaction speed. Kimi describes it as the same model as the standard K2.7 Code, running around 180 Tokens/s in typical coding contexts and up to 260 Tokens/s in short-context cases. CrossModel's current standard detail page is for kimi-k2.7-code.
When to use it
- Long-horizon software engineering: cross-file edits, refactors, DevOps, and debugging that need multi-step planning and tool calls
- Agentic coding: the plan / edit / run-tests / read-logs / self-correct loop inside an IDE or CLI agent
- Multimodal development input: feed screenshots, design mocks, and video clips with text requirements to generate or repair front-end code
- Interactive programming: use the standard model for cost control, and the high-speed tier when turn latency matters more
The CrossModel gateway already serves Kimi K2.7 Code through the OpenAI-compatible /v1/chat/completions interface. Current pricing is in the model catalog.