Int4 g128 symmetric, enable_alg_ext, nsamples=512, seqlen=2048, iters=200, auto-round 0.14.2. Only pin: mtp.fc @ BF16 (lm_head/embeddings/in_proj_a·b BF16 via defaults). 18.7 GB. Full recipe provenance recorded here because it's the metadata we wished every quant shipped.

Measured (2×3090, vLLM v0.25.1, TP=2, 262K, fp8 KV, MTP n=5 — full methodology in club-3090 #662): benchlocal 8-pack 109/150 think-off · 115/150 think-on · MTP accept 56.4% code / 55.1% agent (n=5) · decode 71.9 narr / 106.4 code T/s fresh-boot · NIAH clean to 240K · soak PASS (0/100 silent-empty) · eligible for the int8-activation prefill toggle (VLLM_MARLIN_INPUT_DTYPE=int8, ~+50% prefill, #733).

Serve:

Lineage: layout follows LeaderboardModel1's W4A16 (whose defaults our ablations largely vindicated) with tuned algorithm settings. Build guide + traps (dead-pin regex, export nesting, lm_head loader limits): club-3090 discussions. Follow the base model's license.