Upload README.md with huggingface_hub

README.md

@@ -15,7 +15,9 @@ tags:
 
 NVFP4 (4-bit) quant of [CohereLabs/North-Mini-Code-1.0](https://huggingface.co/CohereLabs/North-Mini-Code-1.0), made with [llm-compressor](https://github.com/vllm-project/llm-compressor) (`nvfp4-pack-quantized`). ~17 GB. Experts quantized; router/gate/`lm_head` kept high-precision.
 
-> Note: HF auto-reports ~17B parameters because it counts the FP4-packed `uint8` tensors (2 weights per byte). The actual model is **30B total / 3B active** (MoE), same as the base model.
+**Benchmarks (DGX Spark / GB10, vLLM):**
+- HumanEval pass@1: **90.2%** — matches the FP8 build (90.2%), within one problem of bf16 (90.9%).
+- Decode: **~57 tok/s** single-request (reasoning off) — **≈1.72×** the FP8 build, at ~17 GB vs ~28 GB.
 
 Serve with vLLM (needs `Cohere2MoeForCausalLM` support + `cohere_melody`):