Roguelite Labs

Kimi K2

426w · 2 min · model, open-weights, coding

Kimi K2

Open-weights model from Moonshot AI (China). Released July 11, 2025. A 1-trillion-parameter Mixture-of-Experts architecture focused on agentic coding and long-horizon tool use. The base lineage for subsequent K2.x versions.

Architecture

  • Total parameters: 1.04 trillion
  • Active parameters per token: 32 billion
  • Experts: 384 (8 selected + 1 shared per token); up from 256 in DeepSeek-V3
  • Attention: Multi-head Latent Attention (MLA)
  • Hidden dimension: 7168 (model), 2048 (MoE expert)
  • Layers: 61
  • Context window: 128K tokens (K2 base); extended to 256K in K2.6
  • License: Modified MIT (allows commercial use)

Benchmarks (Kimi K2, base release)

Benchmark Score
SWE-bench Verified 65.8%
Tau2-Bench 66.1
ACEBench (En) 76.5
LiveCodeBench v6 53.7
AIME 2025 49.5
GPQA-Diamond 75.1

Artificial Analysis Intelligence Index score: 26 (above median of 23 for open-weight non-reasoning models of comparable size at release).

K2.6 (April 2026)

K2.6 is the current production version. Key changes over the base K2:

  • Context window extended to 256K tokens
  • Native vision via a 400M-parameter MoonViT encoder (image and video input)
  • Agent Swarm scales to 300 sub-agents and 4,000 coordinated steps
  • Hallucination rate reduced from 65% (K2.5) to 39%
  • Intelligence Index score: 54 — leading all open-weight models, behind only Anthropic, Google, and OpenAI (all at 57)

K2.6 SWE-bench Pro: 58.6, ahead of GPT-5.4 (57.7), Claude Opus 4.6 at max effort (53.4), and Gemini 3.1 Pro (54.2).

K2.6 pricing: $0.95 / $4.00 per million input/output tokens.

Comparison to Other Open-Weight Models

At the time of the base K2 release, Nemotron Ultra (NVIDIA) was the other prominent open-weight contender at the 1T scale. K2 led on agentic benchmarks (Tau2-Bench, ACEBench) and matched or exceeded it on coding. K2.6 now leads the open-weight category outright on the Artificial Analysis Intelligence Index.

The MoE design — large total parameter count, low active parameters per forward pass — follows the architecture pattern of DeepSeek-V3 but with a larger expert pool.

Availability

Weights available on Hugging Face under Modified MIT License. K2.6 available via Moonshot's API and third-party providers including DeepInfra, Novita, Baseten, Fireworks, and Parasail.

Related

deepseek-r1 · evals · agentic-workflows

Sources