Kimi K2.6

Supported Functionality ItemSpecificationInputText, Image, VideoOutputTextContext262,144 tokens (~256K)Max Output32,768 tokensVision✓ Supported (native multimodal)Function Calling✓ Supported (native agentic) Description Kimi K2.6 is Moonshot AI's flagship open-weight multimodal agentic model, released and open-sourced on April 21, 2026 under a Modified MIT License. Built on a 1-trillion-parameter Mixture-of-Experts architecture with 32B parameters activated per token, it was trained on 15.5T tokens using the Muon optimizer and ships with weights freely available on Hugging Face. It scores 58.6 on SWE-Bench Pro — ahead of GPT-5.4 (57.7) and Claude Opus 4.6 (53.4) — and leads every frontier model on Humanity's Last Exam with tools at 54.0. K2.6's defining breakthrough isn't a single benchmark — it's sustained execution. Where K2.5 could hold a coding task together for a few hundred steps, K2.6 can run autonomously for 13 hours, and its Agent Swarm scales to 300 sub-agents executing across 4,000 coordinated steps simultaneously (up from 100 agents / 1,500 steps in K2.5). This is the first open-source release that turns "agentic coding" from a demo into production infrastructure — at roughly one-fifth the API cost of closed frontier models. Key Capabilities

Long-horizon autonomous coding: Sustains 13-hour unattended coding sessions; internally overhauled an 8-year-old financial matching engine with a 185% median throughput gain. Agent Swarm orchestration: Coordinates 300 sub-agents across 4,000 steps; scores 86.3 on BrowseComp swarm mode, producing entire websites or spreadsheets in one run. Top-tier software engineering: 80.2 on SWE-Bench Verified, 58.6 on SWE-Bench Pro, 89.6 on LiveCodeBench v6, 66.7 on Terminal-Bench 2.0; strong across Rust, Go, and Python. Agentic research: Leads HLE-Full with tools (54.0) over GPT-5.4 and Claude Opus 4.6; achieves 92.5 F1 on DeepSearchQA. Native multimodality: Unified text, image, and video handling in one architecture; powers coding-driven design from visual prompts to production UI. Documents-as-skills: Converts PDFs, spreadsheets, slides, and Word docs into reusable agent skills retaining structural and stylistic DNA. Claw Groups collaboration: Research preview enabling humans and heterogeneous agents on any device/model to collaborate in a shared swarm with K2.6 as adaptive coordinator.

Technical Strengths FeatureBenefitSparse MoE activation1T total params, only 32B active per token — frontier-scale capability at small-model inference cost.Open weightsModified MIT License allows full self-hosting, avoiding data egress concerns and vendor lock-in.Aggressive pricing~$0.95 / $4.00 per million tokens (input/output), roughly 80% cheaper than GPT-5.5 or Claude Opus 4.7.Native INT4 quantizationShips INT4-ready, reducing memory footprint and enabling Mac-local deployment in some configurations.Long-context stabilityRetains full conversation history, tool definitions, and codebases across multi-hour agent sessions at 262K.Ecosystem compatibilityWorks with OpenClaw, Hermes Agent, vLLM, SGLang, and OpenAI-compatible APIs for low-friction migration. Capability Ratings DimensionRatingNotesReasoningExcellentLeads HLE with tools over all major frontier models; top-tier agentic reasoning.CodingTop-tierBeats GPT-5.4 and Claude Opus 4.6 on SWE-Bench Pro; unmatched on long-horizon execution.Creative WritingStrongStable and controllable, though more engineering-leaning than Claude's literary style.MultimodalStrongNative vision and video input with production-grade visual reasoning.Response SpeedFast60–100 tokens/s output throughput thanks to efficient MoE inference.Context WindowHuge262K tokens supports entire codebases and long-running agent histories. Use Cases

Autonomous software engineering: Delivers complete features, migrates legacy systems, and handles multi-day DevOps with minimal supervision. Parallel multi-agent development: Use Agent Swarm to run frontend, backend, test, and documentation sub-tasks concurrently, compressing delivery timelines. Coding-driven design: Generates production-ready frontend interfaces with structured layouts and animations from natural language or mockups in a single pass. Deep research & intelligence: Combines tool use with DeepSearchQA-level retrieval to autonomously produce multi-step cross-document research reports. Enterprise on-prem deployment: Open weights plus INT4 quantization suit finance, government, and other data-sovereignty-sensitive environments. Agent collaboration platforms: Claw Groups lets humans and heterogeneous agents collaborate across devices and model providers in a unified workspace. Document-to-skill workflows: Convert PPTs, PDFs, and spreadsheets into callable agent skills to operationalize organizational knowledge.