Moonshot AI Open-Sources Kimi K2.6 — A Coding Model That Runs Autonomously for Days

Beijing / April 21, 2026 — Moonshot AI has released Kimi K2.6 to the open-source community — a model that executes complex engineering tasks for hours, sometimes days, without a human in the loop. Available immediately via Kimi.com, the developer API, Kimi Code, and Ollama, K2.6 is Moonshot's boldest claim yet that open-source AI has caught up to the proprietary frontier.

Coding That Doesn't Stop at the Prompt

K2.6's defining capability is long-horizon execution — pursuing a real engineering goal autonomously across thousands of steps without stopping to ask for direction.

Two internal cases illustrate how far that goes:

• Mac inference optimization — K2.6 implemented local LLM inference from scratch in Zig, a niche systems language, over 12 hours and 4,000+ tool calls. It pushed throughput from ~15 to ~193 tokens per second — 20% faster than LM Studio.
• Financial engine overhaul — Given an 8-year-old matching engine already near its limits, K2.6 worked for 13 hours, modified 4,000+ lines of code, and extracted a 185% median throughput gain and 133% peak throughput improvement — with no human redirecting it at any point.

What the Benchmarks Show

On the evaluations that matter most for real developer work, K2.6 leads the field:

• SWE-Bench Pro — K2.6 scores 58.6, ahead of GPT-5.4 (57.7), Claude Opus 4.6 (53.4), and Gemini 3.1 Pro (54.2)
• HLE Full with tools — K2.6 leads at 54.0 vs. GPT-5.4 (52.1), Claude Opus 4.6 (53.0), Gemini 3.1 Pro (51.4)
• DeepSearchQA — K2.6 scores 92.5 F1, a clear margin above all three competitors

GPT-5.4 and Gemini 3.1 Pro retain edges on pure reasoning and advanced mathematics. For production engineering work, the gap has effectively closed.

Agent Swarms and Five-Day Autonomous Ops

Swarms at Scale

K2.6's Agent Swarm architecture now supports 300 concurrent sub-agents across 4,000 coordinated steps — up from 100 agents and 1,500 steps in K2.5. In documented tests, the swarm produced 100 tailored resumes from a single CV, generated a full 40-page research paper from one PDF, and built custom landing pages for 30 businesses identified via Google Maps — all in single autonomous runs.

No Human Required

Moonshot's own infrastructure team ran a K2.6-backed agent for five consecutive days, handling system monitoring, incident response, and alert-to-resolution cycles without any human intervention.

What Industry Partners Found

Beta partners were consistent in what they reported:

• Vercel — 50%+ improvement on their Next.js benchmark vs. K2.5
• Factory.ai — +15% on internal benchmarks; fewer shortcuts, better instruction following
• CodeBuddy — +12% code accuracy, +18% long-context stability, 96.6% tool success rate
• Blackbox.ai — surfaces deep bugs that "would normally take significant developer time to uncover"

Read More

• Anthropic Launches Claude Design for Teams to Create Prototypes, Decks, and Landing Pages
• Anthropic Launched Claude Opus 4.5 — New Flagship Model for Coding and Complex AI Workflows
• Microsoft Copilot Adds Multi-Model AI Comparison
• Google Launches Gemma 4, Its Most Capable Open Model Family
• OpenAI Closes Record $122 Billion Funding Round at $852 Billion Valuation