Skip to content

Publication summary: Qwen3.8 27B NInfer NVFP4 MTP3 on RTX 5090

This is derivative publishing copy. The dated finding and linked raw artifacts are authoritative.

Canonical facts

  • Model identity: neroued/Qwen3.8-27B-nvfp4-NInfer revision 204e3d92c30d9d05f3300d2f52e443ad1edf6ddf, artifact SHA-256 bb3360522a06e136e0367f5703414d26272b7285c8a6ab6194135c17dbd81b32
  • Runtime identity: NInfer revision e3aeaf8c0b6f83ae8f051780f0ad0d995d5a7bef, source-built on the digest-pinned CUDA 13.1.2 development image recorded in the recipe
  • Local setup: one RTX 5090, 32,607 MiB; NInfer NVFP4/row-scaled-FP8 target, INT8 KV, MTP3, 252,928 tokens, C1, thinking disabled
  • Matched 4K result: MTP3 held median TTFT near-flat at 0.430 versus 0.421 seconds, raised median decode from 75.3 to 165.9 tok/s, and reduced median E2E from 1.085 to 0.720 seconds across five requests per arm
  • Context result: the MTP3 arm returned the exact marker from a nominal 244,480-token fixture containing 201,746 API-reported prompt tokens while accepting an 8,192-token completion cap; E2E was 70.4 seconds
  • Behavior result: smoke, structured JSON, C1 tools, streaming tools, tool-result continuation, coding 3/3, triage 3/3, and repeated tools 3/3 passed
  • Failed gate: a 20-way shared-prefix tool burst completed 17/20 on both arms; three requests received explicit 429 server_overloaded under the C1 scheduler contract
  • Safety: MTP3 left 2,354 MiB free by nvidia-smi after startup and the long request, above the campaign's explicit 1 GiB model-only floor but below the ordinary 3 GiB reserve
  • Decision: preferred measured RTX 5090 text/tools performance challenger; no-promotion; no route, client catalog, or persistent serve changed

Short post

Qwen3.8 27B NVFP4 + MTP3 on one RTX 5090: 0.43s median TTFT, 165.9 tok/s decode, 0.72s E2E at 4K/C1; 201,746-token prompt + 8,192 output cap passed. Direct challenger only: 20-way burst was 17/20 and no route was promoted.

Claim ledger

Public claim Conditions Evidence
0.430 s TTFT, 165.9 tok/s decode, 0.720 s E2E five 4K-class requests, C1, thinking disabled, cold labeled run capacity-mtp3-4k-c1.json
2.20x matched decode and 33.7% lower median E2E same artifact/runtime/context/KV/C1; MTP3 versus no spec capacity-nospec-4k-c1.json · capacity-mtp3-4k-c1.json
201,746-token prompt with 8,192-token completion cap passed nominal 244,480 fixture, exact marker, thinking disabled preflight-mtp3-needle-244k-output-reserve-8192.json
Repeated bounded coding/triage/tools passed three attempts per deterministic check quality-mtp3.json
Shared-prefix burst was 17/20 20 simultaneous attempts against the C1 recipe; three explicit 429 admissions preflight-mtp3-tools.json
Exact incumbent restored managed health/identity and fresh smoke configuration-end.json · restoration-smoke.json