Open-source model refactors legacy code autonomously for 13 hours 🤌

Moonshot AI dropped Kimi K2.6 into open source.

What's interesting in this release:

1️⃣ Long-Horizon Execution
The blog post described illustrative cases. First: the model was tasked with optimizing local Qwen inference on a Mac. K2.6 made 4000+ tool calls over 12 hours and rewrote the backend in Zig (Zig, can you believe it!), boosting throughput by 20% over LM Studio.
Second: refactoring an 8-year-old financial engine exchange-core. The model autonomously analyzed CPU flame graphs and memory allocations for 13 hours, made over 1000 tool calls, rewrote 4000 lines of code, and changed the thread topology. Result: throughput increased by 185%.

2️⃣ Agent Swarms on steroids
The architecture now scales horizontally to 300 sub-agents that can execute up to 4000 steps in parallel. They decompose the task themselves.

3️⃣ Proactive Agents
Moonshot's internal RL infra team left an agent based on K2.6 for 5 days to autonomously manage monitoring, respond to incidents, and fix the system. Without any human involvement. DevOps engineers, how are you feeling?

🛠 The model beats GPT-5.4 (in xhigh mode) and Claude Opus 4.6 (max effort) on SWE-Bench Pro and HLE (Humanity's Last Exam).
Context window — 262k tokens.
Native INT4 quantization available, runs via vLLM, SGLang, or KTransformers.

Weights available on HF: huggingface.co/moonshotai/Kimi-K2.6