NVIDIA Supercharges Moonshot's Kimi-K2.7-Code to Run 2.35x Faster on Blackwell
NVIDIA quantizes Moonshot AI's 1T-parameter Kimi-K2.7-Code to FP4, making it faster and cheaper to run on Blackwell GPUs without sacrificing accuracy.
Takeaways
- NVIDIA released Kimi-K2.7-Code-NVFP4, a Blackwell-optimized FP4 quantization of Moonshot AI's 1T-parameter coding model.
- NVFP4 delivers 2.35x faster inference than INT4 on Blackwell GPUs with near-zero accuracy loss -- benchmark scores actually improve slightly.
- The base model, Kimi K2.7 Code, cuts thinking-token usage by ~30% vs K2.6 while scoring higher on coding and agentic benchmarks.
- K2.7 Code beats Claude Opus 4.8 on MCP Mark Verified (81.1 vs 76.4), a key benchmark for tool-calling agent workflows.
- Requires NVIDIA Blackwell hardware (B200); served via vLLM with tensor-parallel-size 4; free for commercial use under NVIDIA Open Model License.
- Thinking mode is always on and cannot be disabled -- factor this into token budgeting for agentic pipelines.
NVIDIA just dropped a Blackwell-optimized version of Moonshot AI's Kimi-K2.7-Code-NVFP4 on Hugging Face. The model is a post-training quantized (PTQ) version of Moonshot's latest coding-focused agent model, compressed from its native INT4 format down to NVFP4 -- NVIDIA's 4-bit floating-point format built specifically for Blackwell hardware. The result: a trillion-parameter MoE model that can be served faster, on less memory, with benchmark scores that actually improve over the INT4 baseline.
The model underneath: Kimi K2.7 Code
To understand what NVIDIA shipped, you first need to know what they started with. Kimi K2.7 Code is Moonshot AI's latest open-source coding model, released under a Modified MIT license. It uses a Mixture-of-Experts (MoE) architecture with 1 trillion total parameters and 32 billion active per token. The MoE design means you get the capacity of a 1T model at the compute cost of a 32B dense model -- each token only activates a small subset of the network's 384 experts.
The model is paired with a 256K-token context window. It always operates in thinking mode, preserving full reasoning content across multi-turn conversations. That last point matters for how you budget token spend: K2.7-Code forces thinking mode on, and you can't turn it off. The model always reasons before answering, which affects how you structure workflows and budget token spend.
What K2.7 Code is actually good at
The model ships in a "Code" configuration built for long-horizon software engineering: multi-file refactors, complex debugging, and agentic workflows that run for dozens of turns. The headline benchmark numbers are strong against its predecessor:
- +21.8% on Kimi Code Bench v2 (62.0 vs 50.9), +11.0% on Program Bench (53.6 vs 48.3), and +31.5% on MLS Bench Lite (35.1 vs 26.7) over K2.6.