Kimi K2.7 Code

What it is. Kimi K2.7 Code is Moonshot AI’s dedicated coding model, an API-only member of the Kimi K2.7 family with a 256K-token context window and always-on thinking. Why it matters. It is Moonshot’s code-specialist tier — distinct from the general-purpose, open-weight Kimi K2.6 and the flagship Kimi K3 — priced for high-volume coding and agent work rather than frontier reasoning. Key characteristics. 262,144-token context, reasoning and coding capabilities, $0.95 input / $4.00 output per 1M tokens (cache hits $0.19), and a doubled-price highspeed variant. What a professional should know. Architecture and parameter counts are not disclosed for the K2.7 series, thinking cannot be disabled, and sampling parameters (temperature, top_p, n, penalties) are fixed — you cannot tune them through the API.

Architecture and parameters

Moonshot discloses no architecture or parameter counts for the K2.7 series, and the database records that absence rather than estimating. China AI Hub analysis indicates this is a deliberate API-only posture: Moonshot publishes full architecture cards for its open-weight models (Kimi K2.6, Kimi K3) but keeps the coding line closed and undocumented, which points to a monetization strategy rather than an engineering omission. See Mixture-of-Experts for the architectural context Moonshot does not provide here, and the MoE architectures research for how the disclosure gradient runs across labs.

What the context window actually means

A 262,144-token context is the mid tier of the collection — above MiniMax-M2.7’s 204,800 but a quarter of the 1M windows on Kimi K3, Qwen3.8 and the Doubao Seed Pro/Evolving models. For coding work this is enough to hold a full repository, a large diff, or a long agent trace in a single request, but it is not a long-document-analysis tool. See the context-window research.

Pricing implications

$0.95 input / $4.00 output per 1M tokens with cache hits at $0.19 (20% of input). China AI Hub analysis indicates the 4.2x input-to-output ratio sits between the general K2.6’s identical list price and K3’s 5x flagship ratio, and the aggressive cache-hit discount is aimed squarely at agent loops that re-read shared context. A highspeed variant serves the same model at 2x the price for latency-sensitive work. See the Kimi K3 vs K2.7 Code comparison and the choosing-by-price guide.

API, coding, and agent implications

K2.7 Code targets software engineering: it carries coding and reasoning capabilities and is served through the Moonshot AI Platform. The operational constraints are the story — always-on thinking means every request pays reasoning latency and cost, and fixed sampling (temperature/top_p/n/penalties) removes the usual reproducibility levers. China AI Hub analysis indicates the fixed sampling is a deliberate product choice that trades deterministic-output control for a curated coding response, which is an under-appreciated constraint for batch and CI/CD workloads that expect reproducible generation. The Kimi Code agent is the product built on the K-series. See choosing-a-coding-model and the coding model hub.

Open weights and license

API-only: no open-weight release is confirmed for the K2.7 series, and the model cannot be self-hosted. This contrasts with Moonshot’s open-weight K2.6 (Modified MIT) and K3 (Kimi K3 License). See open weight vs API.

Benchmark interpretation

No benchmark scores are recorded for K2.7 Code in the database. This is an absence of evidence, not evidence of absence — Moonshot has not published K2.7-Code-specific benchmark results on the fetched official pages.

What the benchmarks do not prove

With no recorded benchmarks, there is nothing with which to rank K2.7 Code against competitors, and any third-party claims about its coding ability cannot be verified against official numbers. Treat capability as vendor-positioned until Moonshot publishes results. See how to read vendor-reported benchmarks.

Suitable and less suitable workloads

Well-suited: high-volume coding and agent work on the Moonshot platform where the cache-hit pricing and coding focus pay off, and interactive coding where the highspeed variant’s faster throughput matters. Less suited: long-document analysis (256K, not 1M), workloads needing deterministic or tunable sampling, and open-weight or self-hosting requirements.

Where this model fits

Workload	Relevance
Long-context analysis	Moderate
Coding	High
Structured API workflows	Moderate
Agent orchestration	Moderate
Local self-hosted deployment	No evidence
GUI automation	No evidence
Video generation	No evidence
Enterprise cloud	Moderate

Relevance judgments are China AI Hub analysis based on documented capabilities, not vendor claims.

China AI Hub analysis

China AI Hub analysis indicates that K2.7 Code is Moonshot’s code-monetization line: an API-only, closed-weight coding SKU that sits below the flagship K3 on price ($0.95 vs $3.00 input) while inheriting none of K3’s 1M output headroom or open weights. Its fixed sampling and always-on thinking are the two constraints that most shape real use — together they mean K2.7 Code optimizes for a curated, reasoning-heavy coding output rather than reproducible, tunable generation. The pairing with a 2x-priced highspeed tier is the same throughput-segmentation pattern the database records at MiniMax and Zhipu, which signals that speed-tiered billing is now standard across the Chinese API layer. See the state of China’s AI models research and the reasoning hub.

What is uncertain

• Architecture and parameter counts are not publicly disclosed for the K2.7 series.
• No benchmark scores are published for K2.7 Code.
• The maximum output ceiling is not stated in the fetched docs.
• Sampling parameters are fixed and cannot be modified through the API.

Sources

• Kimi API platform — model list
• Kimi API — pricing (chat)

Labels used above: Official fact (pricing and context window from Moonshot’s docs) and China AI Hub analysis (our synthesis, always introduced as such). No benchmark evidence is currently recorded for this model.