NVFP4 (4-bit) quantization of qwen3-8b-tool-recovery — the QUENCH study's headline artifact: the trained tool-failure-recovery skill is statistically unchanged at 4-bit.

	BF16	this checkpoint (NVFP4)
Recovery Rate (GLM-5.2 judge, 95% CI)	65.6 (57.2–73.9)	65.0 (56.1–73.3) — Δ -0.6pp, ns
Retry behavior	87%	85%
MMLU-400 guard	71.9	68.2 (-3.7pp — knowledge degrades, the trained behavior doesn't)
Size	~16 GB	~6 GB

Quantization: llm-compressor 0.12 NVFP4 scheme (weights+activations), lm_head excluded, calibrated on 512 BFCL-derived chat-templated samples (disjoint from the eval set). Same recipe as the base-model comparison arm — no cross-recipe confound. Note: NVFP4 numbers are recipe-sensitive (a different public NVFP4 recipe scored +9.4pp higher base-model RR); paired conclusions within this recipe are what the study claims.

Serving: requires NVFP4-capable hardware/kernels (Blackwell). Tested with vLLM 0.23 on DGX Spark (GB10):

bash

vllm serve GurkanOz/qwen3-8b-tool-recovery-NVFP4 \
  --enable-auto-tool-choice --tool-call-parser hermes

See the BF16 card for training details, limitations, and the full QUENCH results grid; per-trial transcripts in the dataset.