Model details
Kimi K2.6 continues the Kimi family from Moonshot AI as a frontier-scale open-weight release, scaling to roughly one trillion parameters while retaining the published architecture lineage of the earlier K2.5 generation; the meaningful changes between the two sit in the training recipe rather than the model blueprint, making K2.6 attractive to teams that already operate against K2.5 weights. It is distributed as a truly open-weight model with a Hugging Face license, and is reachable through multiple hosts including the SiliconFlow catalog, where it is positioned as Moonshot AI's latest multimodal agentic offering alongside its prior Kimi-K2-Instruct sibling for direct comparison.
Designed for production agentic workloads, Kimi K2.6 pairs a 262,144-token context window with multi-turn tool calling, vision inputs, and structured outputs, so a single model can read long codebases or document sets, invoke external tools across several turns, and return machine-readable results. The release emphasizes long-horizon coding, proactive autonomous execution, and swarm-based task orchestration, making it a natural fit for coding assistants, multi-step research agents, and automation pipelines that need both deep context and reliable tool use. Available on SiliconFlow with temperature control and open-weight flexibility, K2.6 gives teams a practical path to self-host or consume the same model via API, easing the move from prototyping to deployed agentic systems.
Transparent token rates
Rates are shown per one million tokens. Combined means one million input plus one million output tokens.
SiliconFlow
Kimi K2.6, Moonshot AI's latest open-source multimodal agentic model. This release brings exceptional long-horizon coding capabilities, proactive autonomous execution, and swarm-based task orchestration, enabling developers to achieve production-grade agentic workflows with greater efficiency and reliability.
SiliconFlow
Compare Kimi-K2-Instruct and Kimi-K2.6 across performance, cost, capabilities, and real-world use cases. See which model fits your needs.
SiliconFlow
BenchLM.ai's profile for Kimi K2.6 reports a composite capability score of 59/100, ranking it 66th out of 216 tracked models as of August 7, 2026, with a 256K-token context window, 35 tok/s typical speed, and a 2.94s first-token latency. The model lists API pricing at $0.95 input and $4 output per million tokens. Notab Across eligible benchmark categories, Kimi K2.6 ranks first in Mathematics (1st of 7, 100th percentile) and 55th in Coding (55th of 132, 59th percentile), but lands at 104th of 132 in Agentic tasks (21st percentile) and 48th of 55 in Knowledge (13th percentile), with Multimodal at the 48th percentile. The profile notes
SiliconFlow
SiliconFlow’s June 11, 2026 release notes say that all traffic for moonshotai/Kimi-K2.5 will be automatically routed to moonshotai/Kimi-K2.6 beginning that day. The notes characterize K2.6 as the successor and say it delivers improved performance at the same pricing, with no cost change for users. The release notes recommend that users update model names and migrate to moonshotai/Kimi-K2.6 as early as possible, while noting that Kimi-K2.5 will be deprecated at a future unspecified date. This is direct evidence of a concrete, compatibility-oriented upgrade path for existing SiliconFlow API users.
SiliconFlow
NVIDIA's NGC catalog hosts MoonshotAI Kimi-K2.6 as a downloadable NVIDIA NIM container, confirming the model is open-source and natively multimodal. The architecture combines a Mixture-of-Experts transformer with 1 trillion total parameters and 32 billion activated parameters, paired with a 400M-parameter MoonViT visio Multiple signed NIM container variants are available, including the latest nvfp4-2fd3a800-b3 quantization, b200-config-v6 and h200-config-v4 hardware-specific configurations, an eagle3-d49b3f03-b3 speculative-decoding build, a vlm-nim-with-checksum vision-language container, and an hf-bc8e70-nim Hugging Face-compatible
SiliconFlow
SiliconFlow is listed as one of roughly twenty host providers for Moonshot AI's Kimi K2.6 on OpenRouter's routing page, with input priced at $0.77 per million tokens, output at $3.40 per million tokens, and cache reads at $0.14 per million tokens. SiliconFlow's P50 latency is reported at 3.76 seconds with throughput of The Kimi K2.6 model on this routing page is described as Moonshot AI's next-generation multimodal model with 262K context, targeting long-horizon coding, coding-driven UI/UX generation, and multi-agent orchestration via an agent swarm architecture. OpenRouter lists the model as released April 20, 2026, with promotional
