Skip to content

Publication summary: four efficient Qwen3.8 variants on RTX 5090

Derivative publishing copy only; nothing has been posted externally. The dated finding and native artifacts remain authoritative.

Canonical facts

  • Model identity: Signal27B, Swift27B, Qwopus Flash27B and Minitron20B, exact repositories/revisions/files in recipes; compare the retained Unsloth Qwen3.8-27B GGUF revision 4ca720788d1e01f1bff70c033e0d0028fd02e502.
  • Runtime identity: llama.cpp a298422da78eb75e440a7de0ca408af64d323d93, image digest sha256:cf2e30bc855cf58cdbdc65d05b5b5e02afa95fb788343a5334d704367ac5c9ac.
  • Local setup: one RTX5090,32607MiB, sm120, driver616.64, Windows/Docker/WSL2; challenger Q6_K/Q4_0 KV,64K/C1; incumbent UD-Q4_K_XL/MTP3,262K/C1.
  • Recipe: original four, no-spec controls.
  • Measurement path: direct online loopback. Cache state differs by diagnostic; no matched finalist performance ranking is claimed.
  • Headline result: all four original candidates passed six functional checks, tools20/20 each, and47349 actual prompt-token retrieval.
  • Capability result: ten-question strict thinking-on screen: Signal9/10, Swift9/10, Qwopus1/10, Minitron3/10, incumbent8/10. N1 per question; 5120 combined completion-token cap. This includes format and budget failures.
  • Important caveat: failed128-word output and missing full-context/broad quality/client qualification prevent replacement. Small scores are not broad accuracy.
  • Decision: retain incumbent; no challenger promotion; exact restoration passed. User authority was present, but evidence gates were not cleared.
  • Canonical evidence: finding.
  • Artifact set: manifest, index, summary, restoration.

X / short post

RTX 5090: four Qwen3.8 variants passed tools but none cleared replacement gates. Signal/Swift scored 9/10 on a tiny strict reasoning screen. No promotion. https://fakoli.github.io/anvil-serving/findings/2026-09-12-qwen38-efficient-variants-rtx5090/

Reddit

Four Qwen3.8 variants on one RTX 5090: promising shorter reasoning, but no qualified replacement

I tested Signal, Swift, Qwopus Flash and Minitron on one RTX 5090 using pinned
llama.cpp Q6_K recipes at 64K/C1, including no-spec controls for Signal/Swift.
All four passed the six-check functional gate and tools20/20 each.

On a ten-question strict thinking-enabled sanity fixture, Signal/Swift passed
9/10, versus8/10 for the retained Unsloth UD-Q4_K_XL/MTP3 incumbent. They used
about16% fewer completion tokens. Qwopus and Minitron scored1/10 and3/10;
format and budget failures count, so these are not broad knowledge scores.

No tested profile cleared replacement gates. Exact128-word output failed for
Signal/Swift with and without speculation; revised shorter checks do not erase
that result. The incumbent also failed the original128-word check. The
challengers did not qualify its262K deployment or broader client/quality gates.
I restored the exact incumbent and retained failed attempts. This is a local
diagnostic result, not a universal ranking.

Method, exact recipes and raw evidence:
https://fakoli.github.io/anvil-serving/findings/2026-09-12-qwen38-efficient-variants-rtx5090/

Screenshot alt text

Table of four Qwen3.8 variants tested on a single RTX 5090 with pinned64K/C1 Q6_K recipes. All pass functional tools. Tiny strict thinking-on scores are 9/10 for Signal and Swift,1/10 for Qwopus and3/10 for Minitron, compared with 8/10 for the262K incumbent. Output failures and missing full-contract gates prevent promotion; the incumbent was restored.

Claim ledger

Public claim Conditions Evidence
Four functional passes Six checks and tools20/20 each;47349 actual input tokens Signal, Swift, Qwopus, Minitron
Strict thinking-on scores Ten questions,N1,5120 combined cap; exact syntax and budget failures included Signal, Swift, Qwopus, Minitron, incumbent
About16% fewer completion tokens Total across fixed ten attempts, not isolated reasoning tokens Same five native quality artifacts; finding results
No qualified replacement Original strict failures, tiny diagnostic population and missing full deployment gates Signal failure, Swift failure, summary
Exact incumbent restored Healthy identity, direct functional check, ownership, protected process and shared-memory checks Restoration, preflight, state