Upload README.md with huggingface_hub

README.md

@@ -82,6 +82,23 @@ transformers-5 loading (llmcompressor, finetuning); vLLM serving does not use th
 | [NVFP4-FP8-LoopShield](https://huggingface.co/NullSense/Nanbeige4.2-3B-NVFP4-FP8-LoopShield) | 4-bit family: GSM8K parity with FP8, best JSON, −0.4 GB |
 | [NVFP4A16](https://huggingface.co/NullSense/Nanbeige4.2-3B-NVFP4A16) | speed/summarization: 175/220 tok/s — NOT for reasoning (gates published) |
 
+
+## Full matrix vs the bf16 original
+
+| metric | bf16 original | FP8 (this repo) |
+|---|---|---|
+| freeform tok/s (spec-off) / summarize (+ngram) | 96 / 136 | **154 / 206** |
+| rewrite taxonomy (/20, non-think) | 18 | 18 |
+| blind-judge summarization (same judging pass) | 4.61 | 4.55 (noise-level delta) |
+| JSON parse rate (/48) | 44 | 42 |
+| GSM8K, IFEval, MMLU-Pro, BBH, MultiHop-RAG | not measured on bf16 | measured (tables above) |
+
+Where bf16 was measured, FP8 shows no regression; that is why FP8 serves as the
+reference for the rest of the family instead of re-running every gate on bf16.
+The vendor's published numbers are a different suite entirely (agentic/frontier
+benches, thinking mode at a 131k token budget) and don't map onto these rows —
+see Benchmark provenance below.
+
 ## Links & provenance
 
 - Base model: [Nanbeige/Nanbeige4.2-3B](https://huggingface.co/Nanbeige/Nanbeige4.2-3B)