Integrating Kimi K3 with Claude Code: Running a 2.8T Open-Source Flagship Model via ccmr

Kimi K3 Model Overview

Kimi K3 has 2.8 T parameters and is positioned as the world’s first open 3‑T‑level model. Full weights are scheduled for release by the end of July. Over the past year Kimi held the record for the largest open‑source model size in nine of the last twelve months.

Architecture and Efficiency

The model uses Kimi’s proprietary Delta Attention combined with Attention Residuals. Its Mixture‑of‑Experts layer contains 896 experts, with only 16 activated per token. This design, together with new training and data recipes, yields roughly 2.5× scaling efficiency compared with K2. K3 is natively multimodal, accepts visual inputs, and supports a 1 M token context window.

Performance

Overall performance trails the closed‑source leaders Claude Fable 5 and GPT 5.6 Sol, but K3 ranked first in the Arena front‑end‑code benchmark, pushing Fable 5 to second place.

Pricing

Cache‑hit input cost is $0.30 per million tokens, cache‑miss input cost $3.00 per million tokens, and output cost $15.00 per million tokens.

Claude Code Model Router (ccmr)

ccmr (Claude Code Model Router) is a lightweight gateway released in December 2023 that enables Claude Code to run on domestic models without an official Claude subscription. GitHub repository: https://github.com/luwill/Claude-Code-Model-Router

Versions 1.9.0 and 1.10.0 add three entry points for K3:

International API short name kimi-k3 (alias k3) using the existing KIMI_API_KEY.

Domestic open‑platform short name kimi-cn-k3 requiring a separate KIMI_CN_API_KEY; the platform domain changed from platform.moonshot.cn to platform.kimi.com.

Kimi Coding Plan users: version 1.10.0 introduces a subscription entry using a KIMI_CODE_API_KEY obtained from the Kimi member console. Subscription tier determines the context window size: kimi-plan-k3-1m (Allegretto tier, 1 M window) or kimi-plan-k3 (Moderato tier, 256 K window).

Typical Usage Commands

# Install ccmr
npm install -g claude-code-model-router

# Initialize configuration (adds .env to .gitignore)
ccmr init            # or: ccmr init --global  # writes to ~/.ccmr

# Optional connectivity check
ccmr doctor

# Launch Claude Code (gateway starts automatically if not running)
ccmr claude

# Switch default model
ccmr use kimi-plan-k3

After setting K3 as the default model with ccmr use k3, invoking ccmr claude runs K3. In the author’s tests K3 responded aggressively, generating a full workflow for even trivial queries.