Kimi K2.7-Code suits teams building self-hosted or budget-conscious agentic coding tools who are comfortable running their own evaluations rather than trusting vendor benchmarks. It uses roughly 30% fewer thinking tokens than its predecessor on equivalent tasks, which Moonshot says lowers effective cost. Regulated teams needing compliance documentation should look elsewhere.
Kimi K2.7-Code posts a 21.8% gain over Moonshot's prior release on the vendor's own Kimi Code Bench v2, the sharpest jump yet in the Kimi K2 family. It is an open-weight, coding-specialized Mixture-of-Experts language model built for long-horizon software engineering: codebase review, debugging, and multi-step agentic tool-calling, released under a Modified MIT license that permits free self-hosting.
Where it sits
- $1.71/M$ per 1M tokensBlended price (3:1)Lower is better#33 / 64peer median $1.70/Mvendor price, checked by HokAI
- --tokens/sOutput speedHigher is better-- / 39peer median 90 tok/scited: Artificial Analysis
- --% solvedSWE-bench VerifiedHigher is better-- / 28peer median 78.3%per source, see benchmark scores
- --% correctGPQA DiamondHigher is better-- / 44peer median 88.3%per source, see benchmark scores
Priced around the middle of the 64 GA models with a published price (rank 33), and one of 4 whose vendor states it may train on customer data.
Ranks are against GA models on HokAI that publish the same figure; ties share a rank.
Provider: Moonshot AI · Family: Kimi K2
More about Moonshot AI on HokAI
Context window: 262,144 tokens
Input modalities: text, image, video, tool-calls, code · Output: text, tool-calls, code
About Kimi K2.7-Code
Kimi K2.7-Code is Moonshot AI's coding-specialist release, launched June 12, 2026 as the fifth major update in the Kimi K2 family in under a year. It shares its Mixture-of-Experts architecture, 384 experts routed 8 per token via Multi-head Latent Attention, with the prior Kimi K2.5 and Kimi K2.6 releases, but is fine-tuned specifically for long-horizon software engineering: codebase analysis, debugging, refactoring, and tool-calling inside agentic coding loops.
Moonshot has shipped a new K2-family model roughly every two to three months since the original Kimi K2 launched in July 2025, through K2 Thinking, K2.5, K2.6, and now K2.7-Code, and the cadence suggests another release within a similar window. K2.7-Code reuses its predecessors' architecture rather than introducing a new base model, so this release reads as a fine-tuning and data update aimed at coding workloads specifically, not a general-capability upgrade.
Pricing
Moonshot's official API charges $0.95 per million input tokens ($0.19 on a cache hit) and $4.00 per million output tokens, matching the prior release's rate. OpenRouter routes the same model through a third party at roughly $0.75 per million input tokens and $3.50 per million output tokens. The Modified MIT license also permits free self-hosting, so the only cost for teams running their own infrastructure is compute.
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.028 | $0.0040 | $0.033 |
| Support reply | $0.0019 | $0.0012 | $0.0031 |
| One coding agent run | $0.190 | $0.080 | $0.270 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Key Features
- Mixture-of-Experts Architecture: Combines 384 experts, 8 routed per token, with Multi-head Latent Attention, the same base design used across the two prior Kimi K2 releases.
- 256K Context Window: Handles up to 256K tokens of codebase, transcript, or diff in a single call, the same window size as the two prior K2 releases.
- Always-On Interleaved Thinking: Runs in a forced reasoning mode with state preserved across turns, which Moonshot says cuts thinking-token usage by roughly 30% compared to its predecessor.
- Native Multimodal Input: Accepts text, image, and video input through Moonshot's vision encoder, useful for coding tasks that involve screenshots, diagrams, or UI mockups.
- Modified MIT Open Weights: Full weights are downloadable from Hugging Face for self-hosting and commercial use, unlike closed frontier coding models that require an API subscription.
Pros
- Reports a 21.8% gain on Moonshot's own Kimi Code Bench v2 over its prior release, the largest coding-specific jump in the K2 series so far.
- Open weights under a Modified MIT license let teams self-host the full model for free, unlike closed competitors that require a paid API.
- Supports MCP-based tool use with reasoning state preserved across calls, which suits long multi-step agent sessions better than models that reset context each turn.
Cons
- No independent third-party benchmark scores exist yet: every published number, including Kimi Code Bench v2, comes from Moonshot's own proprietary suite.
- No safety system card, red-team disclosure, or compliance certification (SOC 2, GDPR, HIPAA) has been published, and the hosted API processes data in China with no training opt-out.
- Self-hosting the full weights needs about 340GB of storage and 64GB or more of VRAM for FP16 inference, ruling out single-consumer-GPU setups.
Benchmarks
- Program Bench Delta Pct Vs K2 6: 11 vendor-reported · 12 Jun 2026
- Mls Bench Lite Delta Pct Vs K2 6: 31.5 vendor-reported · 12 Jun 2026
- Kimi Code Bench V2 Delta Pct Vs K2 6: 21.8 vendor-reported · 12 Jun 2026
- Thinking Token Reduction Pct Vs K2 6: 30 vendor-reported · 12 Jun 2026
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Frequently Asked Questions
How much does Kimi K2.7-Code cost in 2026?
Moonshot's official API prices this model at $0.95 for every million input tokens, or $0.19 apiece on a cache hit, plus $4.00 for every million output tokens, the same headline rate as the prior K2.6 release. OpenRouter's third-party listing charges about $0.75 per million input tokens and $3.50 for a million output tokens instead. Because the model is released under a Modified MIT license, teams can also self-host the open weights and pay only for their own compute, with no per-token fee.
Is Kimi K2.7-Code free to use?
There's no free tier on Moonshot's hosted API for K2.7-Code. The open weights are released under a Modified MIT license, though, so you can download and run the full model yourself for the cost of your own compute, with no license fee. Cloudflare Workers AI and Vercel's AI Gateway both offer pay-as-you-go access if you'd rather not manage your own infrastructure.
What are the best alternatives to Kimi K2.7-Code?
DeepSeek V4 Pro is the closest open-weight alternative: it publishes an 80.6% SWE-bench Verified result and offers a 1M-token context window under an MIT license. Qwen3.8-Max is a proprietary option from Alibaba with a larger context window if you don't need self-hostable weights. For teams that need independently verified frontier benchmarks, better-documented options exist from Anthropic and OpenAI, though at a higher per-token price.
Is Kimi K2.7-Code better than DeepSeek V4 Pro?
It depends on what you're optimizing for. DeepSeek V4 Pro has a larger 1M-token context window and a published SWE-bench Verified score, while Kimi K2.7-Code has a cheaper input-token price and uses fewer thinking tokens per task, which Moonshot says lowers the effective cost per completed job. Neither model has been benchmarked head-to-head by an independent third party as of mid-2026.
How long does it take to get going with Kimi K2.7-Code?
The fastest path is Moonshot's OpenAI-compatible API: create an account, generate a key, and point an existing OpenAI SDK client at the new base URL, which takes a few minutes. OpenRouter and Cloudflare Workers AI offer the same model through their own dashboards if you already use those platforms. Self-hosting takes longer: the weights are on Hugging Face, but downloading roughly 340GB and provisioning 64GB or more of VRAM (or a quantized build for smaller hardware) is a multi-hour setup, not a five-minute one.
Top Alternatives
- DeepSeek V4 Pro: Pick DeepSeek V4 Pro if you want a published 80.6% SWE-bench Verified score and a 1M-token context window; pick Kimi K2.7-Code if you want a lower input-token price and don't need the extra context.
- Qwen3.8-Max: Pick Qwen3.8-Max if you want Alibaba's larger proprietary model with a published benchmark score; pick Kimi K2.7-Code if you need open weights you can self-host under a Modified MIT license.
- Kimi K3: Pick Kimi K3 if you need Moonshot's newer, larger flagship and a wider context window; go with Kimi K2.7-Code instead if you specifically want the cheaper, coding-tuned option on the smaller architecture.
HokAI guides covering Kimi K2.7-Code
- ChatGPT vs Kimi: Which Should You Use in 2026?: ChatGPT costs more but does more. Kimi scores higher on reasoning benchmarks and costs less per token, but its paid plans are sold out. The 2026 verdict.