Kimi K2.6 Goes Open Source: 1T MoE Architecture, Coding Benchmarks Near GPT-5.4 and Opus 4.7

Kimi K2.6 Goes Open Source: 1T MoE Architecture, Coding Benchmarks Near GPT-5.4 and Opus 4.7

Moonshot AI open-sourced Kimi K2.6 in late April. This is not a routine model iteration—it approaches the level of top closed-source models on multiple benchmarks while remaining fully open-source.

Core Specs

  • Architecture: 1T parameter MoE (Mixture of Experts)
  • Context Window: 256K tokens (training platform supports 265K)
  • License: Open weights, available through Nous Portal, Cline, Fireworks AI
  • GitHub: MoonshotAI/Kimi-K2 — 10,700 stars

Benchmark Performance

BenchmarkKimi K2.6Reference
LiveBenchSurpasses Opus 4.7Highest among open models
Terminal-BenchNear GPT-5.4, Opus 4.7~1/6 the cost
Document Arena#8, +14 pts over K2.5-Thinking#1 open model
Vision Arena#15, +9 pts over K2.5-Thinking#1 open vision model

Agent Capabilities

  • Supports 300 parallel sub-agents from a single prompt
  • Can execute 4,000 coordinated steps
  • Supports 12 hours of autonomous operation
  • Covers coding, research, slides, spreadsheets, dataset generation, document writing

Available Channels

  • Nous Portal: Free trial (powered by Vercel AI Gateway)
  • Cline: Limited-time free access
  • Fireworks AI: SFT, DPO, RL fine-tuning support, 265K context
  • Cloudflare Workers: Direct deployment
  • Hugging Face: Open weights download

Quick Start

from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("MoonshotAI/Kimi-K2.6", trust_remote_code=True)

Key Sources