Moonshot AI has released Kimi K2.6, a 1 trillion-parameter vision-language model aimed at extended autonomous coding sessions and larger-scale multi-agent orchestration. The company says the model performs competitively with open-weights peers such as Qwen3.6 Max Preview and the newly released DeepSeek V4, while remaining slightly behind leading closed-source models.
Capabilities of Kimi K2.6
- Architecture: mixture-of-experts design with a total of 1 trillion parameters and 32 billion active parameters per token. Visual encoding is handled by the MoonViT encoder (400 million parameters). Kimi K2.6 builds on the architecture introduced in Kimi K2 and refined in Kimi K2.5, including multi-headed latent attention.
- Input/output: accepts text, images, and video up to 256,000 tokens and can generate text outputs up to 98,000 tokens.
- Quantization: trained with native INT4 quantization.
- Features: tool use, web search, a "preserve thinking" mode that keeps earlier reasoning tokens across multi-turn interactions, and an agent swarm mode for coordinated multi-agent work.
- Agent swarm specifics: a coordinator agent decomposes tasks into subtasks and can spawn up to 300 parallel subagents, each able to execute 4,000 steps (up from 100 subagents and 1,500 steps in Kimi K2.5). Work is reassigned when an agent fails or stalls. A research-preview feature called "claw groups" lets agents from other developers — which can run on any device or model — and human collaborators join the swarm.
Performance and benchmarks
- Moonshot reports that Kimi K2.6 leads many open-weights models on intelligence and agent-capability benchmarks and scores well in subjective human-preference tests, but some top proprietary models outperform it on reasoning and large-scale coding benchmarks.
- Artificial Analysis Intelligence Index: Kimi K2.6 scored 54 on the reasoning setting, leading open-weights models but trailing GPT-5.5 set to xhigh reasoning (60) as well as Claude Opus 4.7 (max reasoning) and Gemini 3.1 Pro Preview (reasoning). The closest open-weights competitors, Qwen3.6 Preview and DeepSeek-V4-Pro (both set to max reasoning), tied at 52.
- Strengths: Kimi K2.6 ranked highly on GPQA Diamond (graduate-level science Q&A), HLE (expert-level multidisciplinary reasoning), and SciCode (scientific-code generation). It was, however, narrowly behind DeepSeek-V4-Pro on five index benchmarks and underperformed Xiaomi MiMo-2.5-Pro and other open-weights models on two tests.
- Large-scale coding test: Moonshot tasked Kimi K2.6 with porting Qwen3.5-0.8B's inference code to Zig and optimizing it for a Mac. Over 4,000+ tool calls and 14 successive revisions across more than 12 hours, Kimi K2.6 increased throughput from roughly 15 to 193 tokens per second and finished about 20% faster than LM Studio running on the same hardware.
- Hallucination rate: Artificial Analysis measured a hallucination rate of 39.26% on a general-knowledge QA benchmark (the share of incorrect outputs that include erroneous answers, false claims of ignorance, or refusals). This is lower than Kimi K2.5's 64.6% and roughly comparable to Anthropic Claude Opus 4.7 (36.18%).
- Arena.ai Code Arena WebDev leaderboard: as of April 26, 2026, Kimi K2.6 had 1,529 Elo and ranked sixth out of 67 models; ahead were Anthropic Claude Opus 4.7 (1,565 Elo), Claude Opus 4.6 (1,548 Elo), and Z.ai's open-weights GLM-5.1 (1,534 Elo).
Availability, license and pricing
- Weights: Kimi K2.6 weights are available for download on Hugging Face under a modified MIT license that permits commercial use but requires attribution for products with more than 100 million monthly active users or over $20 million in monthly revenue.
- Interfaces and API: Moonshot offers a free chat interface at kimi.com, a Kimi mobile app, and API access priced at $0.95/$0.16/$4.00 per million input/cached/output tokens.
Undisclosed information
Moonshot has not disclosed the detailed training data or the specific training methods that differentiate Kimi K2.6 from Kimi K2.5.
Context and why it matters
Long-horizon autonomous execution — the ability of an agent to remain on task across many hours or days — became a competitive focus in late 2025. Anthropic, OpenAI, and Alibaba all targeted this capability in recent releases. Kimi K2, released in July 2025, was an early open-weights entrant for agentic tool use; the family has seen multiple updates with increasing emphasis on sustained execution.
By extending how long Kimi models can autonomously carry out useful work — from short reasoning traces to multi-step tool use, multi-hour coding, and now multi-day projects — Moonshot reduces the frequency of human check-ins needed to keep agents on track.
Brief note
Moonshot highlights that sustained autonomy and lower hallucination rates are related: an agent that can check and revise its own work can detect and correct mistakes, which is central to the design goals behind Kimi K2.6.
