Skip to content

Publication summary: Qwen3.8 27B quant bakeoff on RTX 5090

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

Canonical facts

  • Hardware: one RTX 5090, 32,607 MiB, isolated direct managed recipe lane
  • Reserve: zero added reserve only after a process-free idle baseline; this was a campaign exception, not a policy change
  • Primary metric winner: Gittensor target-only NVFP4/SGLang, 50.9 ms warm median TTFT and 79.5 tok/s median decode at a 3,613-token prompt
  • Capacity: exact retrieval passed with 244,002 API-reported prompt tokens
  • Decode winner: CometKim NInfer MTP3 at 228.0 tok/s, disqualified as a general-purpose winner by a 0/3 strict tool-schema result
  • Balanced full-context fallback: cdiamond iMatrix GGUF MTP8, 223.1 ms TTFT and 96.0 tok/s decode
  • Newest Unsloth arm: Dynamic V3.0 NVFP4 MTP3, 388.7 ms TTFT, 137.7 tok/s short decode, 127.5 tok/s at 53,706 prompt tokens, tools 20/20
  • Speculation failure: Gittensor's advertised DSpark pair failed CUDA- graph capture on incompatible matrix shapes, not memory
  • Quality caveat: Gittensor SGLang used default 1.0 FP8 KV scales because calibrated scales were absent
  • Restoration: exact Unsloth GGUF incumbent restored; smoke, JSON, and 20/20 shared-prefix tools passed
  • Decision: direct TTFT challenger, no-promotion; no route or client configuration changed

Short post

Qwen3.8 27B on one RTX 5090: Gittensor target-only NVFP4/SGLang won TTFT at 50.9 ms median and passed a 244,002-token prompt; CometKim MTP3 won decode at 228 tok/s but failed strict tools 0/3. Incumbent restored, no promotion.

Claim ledger

Public claim Conditions Evidence
50.9 ms median TTFT and 79.5 tok/s decode five warm C1 requests, 3,613 prompt tokens, thinking disabled gittensor-nospec-timing-4k-c1-warm.json
244,002 actual-prompt retrieval pass 252,928 planned context fixture, exact marker gittensor-nospec-capacity-252928-c1.json
Gittensor bounded coding/triage/tools passed three attempts per deterministic check plus 8K/32K context gittensor-nospec-quality.json
CometKim MTP3 strict tools failed 0/3 expected function selected; required zip string absent cometkim-mtp3-quality.json
Unsloth Dynamic V3 MTP3 reached 137.7 tok/s five warm C1 requests, pinned stock vLLM, 64K profile unsloth-nvfp4-mtp3-timing-4k-c1-warm.json
Unsloth tools 20/20 and 53,706-token prompt passed thinking disabled, MTP3, FP8 KV unsloth-nvfp4-mtp3-preflight.json · unsloth-nvfp4-mtp3-capacity-57344-c1.json
DSpark failed on matrix shapes exact publisher target/draft/runtime profile gittensor-dspark-load-failure.json
Dedicated idle baseline had no GPU process incumbent unloaded before reserve decision gpu-idle-baseline.json
Exact incumbent restored identity, managed health, and fresh functional checks restoration.json · restoration-preflight.json