Kimi K2.6

2026-04-20 · Publisher : Moonshot AI · Domain : Artificial Intelligence

Moonshot AI publie Kimi K2.6, modèle ouvert multimodal de 1000 milliards de paramètres axé sur le codage de longue durée et un essaim pouvant coordonner 300 sous-agents sur 4000 étapes.

Moonshot AI releases Kimi K2.6, a multimodal open model with 1 trillion parameters focused on long-horizon coding and a swarm able to coordinate 300 sub-agents over 4,000 steps. ### Context In April 2026, GPT-5.4, Claude Opus 4.6 and Gemini 3.1 Pro are the closed references. Kimi K2.6 succeeds K2.5 (January 2026) with the same 1T-parameter base, but targets reliability over long coding sessions and agent autonomy. ### What the model brings - Long-horizon coding: published example of deploying and optimizing an inference engine in Zig, taken from about 15 to 193 tokens per second over more than 4,000 tool calls and 12 hours. - Expanded Agent Swarm: 300 sub-agents and 4,000 coordinated steps (versus 100 and 1,500 for K2.5). - Generation of complete interfaces from a prompt, with lightweight full-stack flows. - Proactive agents 24/7 and “Claw Groups” (research preview) for collaboration between heterogeneous agents. - Native INT4 quantization carried over from K2 Thinking. ### Architecture and specifications MoE with 1T parameters, 32B active, 384 experts, 256K context (262,144 tokens), MoonViT encoder, text, image and video input, thinking and instant modes. License: modified MIT (display of “Kimi K2.6” beyond 100 million monthly active users or 20 million dollars in monthly revenue). Weights: huggingface.co/moonshotai/Kimi-K2.6. API identifier: kimi-k2.6. Available on Kimi.ai, the app, the API and Kimi Code. ### Cost and access | Item | Pricing (USD per million tokens) | |---|---| | Input (cache miss) | 0.95 | | Input (cache hit) | 0.16 | | Output | 4.00 | Pricing from the official page consulted on September 29, 2026; the pricing on launch day could not be verified. Hardware: not disclosed. ### Benchmarks Scores published by the vendor (Hugging Face card). The asterisks in the source flag re-evaluations by the vendor for the closed models. | Benchmark | K2.6 | GPT-5.4 (xhigh) | Claude Opus 4.6 | Gemini 3.1 Pro | |---|---|---|---|---| | HLE-Full, with tools | 54.0 | 52.1 | 53.0 | 51.4 | | SWE-Bench Pro | 58.6 | 57.7 | 53.4 | 54.2 | | SWE-Bench Verified | 80.2 | not disclosed | 80.8 | 80.6 | | Terminal-Bench 2.0 | 66.7 | 65.4 | 65.4 | 68.5 | | BrowseComp | 83.2 | 82.7 | 83.7 | 85.9 | | GPQA-Diamond | 90.5 | 92.8 | 91.3 | 94.3 | | AIME 2026 | 96.4 | 99.2 | 96.7 | 98.3 | Source: Kimi-K2.6 Hugging Face card (vendor measurements). ### Versus closed models According to the vendor, K2.6 is ahead of the three closed models cited on HLE with tools (54.0) and on SWE-Bench Pro (58.6 versus 57.7, 53.4 and 54.2). On SWE-Bench Verified it is 0.6 points behind Claude Opus 4.6 and 0.4 behind Gemini 3.1 Pro. It remains behind in pure reasoning: GPQA-Diamond 90.5 versus 92.8 to 94.3, and AIME 2026 96.4 versus 96.7 to 99.2. The API price (0.95 and 4.00 USD) is published, but no pricing for the closed models cited was consulted: the price gap is not quantified. The open-versus-closed choice therefore becomes mostly a question of sovereignty and licensing rather than a clear trade-off in level on agentic coding. ### Key takeaways Kimi K2.6 is the point where an open model shows parity with the previous generation's closed models in agentic coding, according to the vendor's measurements. It remains behind in pure reasoning.

Tags : LLM, Open Source, Moonshot AI, Kimi, MoE, MIT modifiée, Codage, Agentique