Moonshot AI Open-Sources Kimi K2.7-Code, a Trillion-Parameter Coding Model That Cuts Reasoning Tokens by 30%

Overview

Moonshot AI has released Kimi K2.7-Code, an open-weight coding-focused large language model, publishing the weights on Hugging Face under a Modified MIT License. The release continues a rapid cadence from the Beijing lab, which has shipped a string of trillion-parameter Kimi models over the past year. The defining claim of the new model is efficiency: Moonshot says K2.7-Code reduces “thinking-token usage by approximately 30% compared with Kimi K2.6”, the metric most directly tied to inference cost for teams running agentic coding workflows.

What We Know

According to the model card on Hugging Face, Kimi K2.7-Code uses a mixture-of-experts (MoE) architecture with 1 trillion total parameters and 32 billion activated parameters per token. The card lists 384 experts, of which 8 are selected per token alongside 1 shared expert, across 61 layers, and a context length of 256K tokens. The same architecture footprint — 384 experts with 8 selected per response — was described for the predecessor K2.6, which Moonshot released in April, according to SiliconANGLE.

The model is positioned as an agentic system for long-horizon coding tasks, with support for multi-step tool calling and interleaved thinking, per the model card. Moonshot notes that the configuration forces reasoning to remain on by default, a design choice the company pairs with its claim of lower overall token consumption.

Moonshot reports double-digit gains over K2.6 on its own coding benchmarks. The company says K2.7-Code improved by 21.8% on Kimi Code Bench v2 and 31.5% on MLS Bench Lite, as reported by VentureBeat. On the model card’s published figures, Kimi K2.7-Code scores 62.0 on Kimi Code Bench v2 and 76.0 on MCP Atlas, trailing the closed-weight GPT-5.5 (69.0 and 79.4) and Claude Opus 4.8 (67.4 and 81.3) on those two tests, according to the Hugging Face card.

The release extends Moonshot’s open-weight strategy. The company made the earlier Kimi K2.5 — also a 1-trillion-parameter MoE model — available on Hugging Face in January, and it has continued to publish weights for subsequent versions rather than gating them behind a proprietary API. Moonshot’s open-weight Kimi line was also the basis for its valuation surge earlier this year, when the lab closed a $2 billion round at a $20 billion valuation.

What We Don’t Know

The performance claims rest largely on benchmarks that Moonshot itself runs. Kimi Code Bench v2, Program Bench and MLS Bench Lite are proprietary tests, and VentureBeat reported that practitioners questioned whether the headline gains hold up in independent use, under the headline “Kimi K2.7-Code cuts thinking tokens 30% — but practitioners say the benchmarks don’t check out.” The 30% reduction in reasoning tokens is likewise a vendor-reported figure that independent testing has not yet confirmed. The model card’s comparison table also shows K2.7-Code still trailing the leading closed-weight systems on the two benchmarks it publishes head-to-head numbers for, leaving open how the open-weight model compares across a broader, independently run suite.

Analysis

K2.7-Code reflects a now-familiar pattern in the open-weight segment: rather than chasing a higher ceiling on raw capability, the update is framed around cost. By cutting reasoning-token usage while keeping the same trillion-parameter MoE backbone as K2.6, Moonshot is targeting the operational economics of agentic coding, where token volume — not just accuracy — determines whether a model is viable in production. Whether the efficiency gains translate outside Moonshot’s own benchmarks is the open question independent evaluators will now test.