Kimi K2.7 from Moonshot: An Affordable AI Model Against GPT-5.5

Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisis ton rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
Moonshot AI recently unveiled Kimi K2.7 Code, an open-source artificial intelligence model that stands out for its competitive pricing and technical capabilities. Designed for complex programming tasks and agent-based workflows, Kimi K2.7 Code is priced at $0.95 per million input tokens and $4.00 per million output tokens. This pricing strategy makes it an attractive option, especially compared to models like GPT-5.5 and Claude Opus 4.8, whose costs are significantly higher, reaching $5.00 and $30.00 per million input and output tokens, respectively.
Although Kimi K2.7 Code does not compete with its Western counterparts on all coding benchmarks, it shows a significant improvement over its predecessor, Kimi K2.6. On Moonshot's internal testing suite, Kimi Code Bench v2, its performance increased from 50.9 to 62.0. On Program Bench, it progressed from 48.3 to 53.6, and on MLS Bench Lite, from 26.7 to 35.1. In agent-focused tests, K2.7 Code achieved 76.0 on MCP Atlas and 81.1 on MCPMark Verified, surpassing Claude Opus 4.8 in the latter test with a score of 76.4.
Kimi K2.7 Code is based on a Mixture-of-Experts (MoE) architecture with one trillion parameters, of which only 32 billion are active per token. This multimodal model can process images and videos, thanks to a custom vision encoder, MoonViT, which has 400 million parameters. One notable improvement is its reasoning efficiency, using 30% fewer reflection tokens than K2.6. The model enforces a reflection mode and a "preserve_thinking" mode that retains complete reasoning content over multiple conversation turns to enhance performance in agent-based coding scenarios.
Moonshot AI plans to introduce a "high-speed mode 6x" for K2.7 Code, which will be accessible via the Kimi API, Kimi Code CLI, and inference engines like vLLM and SGLang. The model is also available in native INT4 quantization, facilitating its execution on less powerful hardware. The model weights are downloadable on Hugging Face, allowing for greater usage flexibility.
Distributed under a modified MIT license, Kimi K2.7 Code allows for free use but imposes increased visibility for large commercial users. Anyone using K2.7 Code or its derivatives in commercial products with more than 100 million monthly active users or generating over $20 million in monthly revenue must display "Kimi K2.7 Code" prominently in the user interface. This pricing and licensing strategy could redefine competitiveness in the AI model space, where cost per token becomes as crucial a criterion as raw performance.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.