Moonshot AI Open-Sources Kimi K2.6: 13 Hours of Uninterrupted Coding, SWE-Bench Surpasses GPT-5.4

Moonshot AI Open-Sources Kimi K2.6: 13 Hours of Uninterrupted Coding, SWE-Bench Surpasses GPT-5.4

On the evening of April 20, Moonshot AI released and open-sourced Kimi K2.6 under the Apache 2.0 license — a programming large model with 1 trillion parameters, hailed as “the most powerful open-source code model to date.”

Core Capabilities

  • 13 Hours of Uninterrupted Coding: Supports continuously writing or modifying over 4,000 lines of code, covering Python, Rust, Go, and more
  • Autonomous Agent Execution: Given complex instructions like “generate a 2026 AI coding assistant report,” it can automatically break down tasks, execute step by step, and self-verify
  • Agent Swarm Upgrade: Supports multiple agents collaborating to complete larger-scale coding tasks

Benchmarks

On core programming benchmarks such as SWE-Bench Pro and Terminal-Bench 2.0, K2.6 surpassed GPT-5.4 and Claude Sonnet, setting a new record for open-source models.

Pricing Strategy

API prices were raised by 58%, officially explained as “a reasonable alignment of compute costs and model scale.” It delivers programming performance close to flagship models at 1/8 the price.

Open-Source Threat

Moonshot AI adopts a “open-source models, charge for services” strategy, consistent with DeepSeek’s approach. For OpenAI and Anthropic, the open-sourcing of trillion-parameter programming models poses a greater threat than any paid competitor — because the pricing power of closed-source models is fundamentally challenged.

Shortcomings

Long-horizon reasoning remains insufficient; multi-step logical deduction for math problems is not as stable as coding tasks.


Main sources: Toutiao, Baidu Baike, 21st Century Business Herald