halt95/Qwen3.8-Flash-Next-W4A16-Merlin

model card: restore the measured three full 262K sessions (v2); v2.2 re-tested at 3 x 256K

model card: v2.2.0 corrections (quality metric, ~256K residency, caveats, metadata)

model card: v2.2.0 (vLLM 0.30 runtime, greedy repeatability, any-rank fail-closed)

model card: v2.2.0 (vLLM 0.30 runtime, greedy repeatability, any-rank fail-closed)

Update README.md

model card: v2.0.1 (806K KV pool, TP2xPP2, structured output under concurrency fixed, known behaviours)

model card: v2.0.1 (806K KV pool, TP2xPP2, structured output under concurrency fixed, known behaviours)

model card: v2.0.1 (806K KV pool, TP2xPP2, structured output under concurrency fixed, known behaviours)

model card: v2.0.1 (806K KV pool, TP2xPP2, structured output under concurrency fixed, known behaviours)

model card: v2.0.1 (806K KV pool, TP2xPP2, structured output under concurrency fixed, known behaviours)

v2: ple_embedding_dtype in text_config (the one v2 config delta)

card: audit round 3 fixes; index metadata total_shards 17 -> 27

serving figures from the 2026-09-08 bench card (shipped sidecar)

model card: all four cards Gen4 x16

model card: host hardware named

model card: allocator numbers and prefix-cache provenance aligned with the code release

Qwen3.8-Flash-Next-W4A16-Merlin weights

model card: corrections from two independent truth checks against the run records

model card: Ampere rationale first, lineage corrected

Qwen3.8-Flash-Next-W4A16-Merlin weights

calibrated fp8 KV scale sidecar (2026-09-08)

Qwen Community License 1.0 (vendored from the base model)

model card

initial commit