Skip to content

Publication summary: Qwen3.8 27B NVFP4 RTX 5090 128K result

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

Canonical facts

  • Model identity: RadixArk/Qwen3.8-27B-NVFP4@554ebba9b5f1b79dc11246341960360e6ef05ef4, served as qwen38-27b-radixark-nvfp4-sglang-rtx5090-128k-mm
  • Runtime identity: SGLang c4271c3fe1262fc2adbd162c33b25de5255251c5, image sha256:506525a5907ea22c9d445afb7c03603959b912de034d86915cf17da814f1a124
  • Local setup: one NVIDIA GeForce RTX 5090, 32,607 MiB, sm_120; isolated Windows 11 / Docker Desktop / WSL2 lane; ModelOpt NVFP4/FP8 weights, FP8 E4M3 KV, 131,072 context, c1, MTP and thinking disabled
  • Recipe: managed 128K multimodal recipe at 85b21147
  • Measurement path: direct online managed endpoint; retained artifacts do not classify individual requests as cold or warm
  • Headline result: exact retrieval marker at 119,675 prompt tokens plus 14 completion tokens in 29.811 seconds; tool calls passed 20/20
  • Capability result: established multimodal corpus passed 30/30 and the corrected eight-image/two-video boundary corpus passed 4/4; direct image/mixed/video latency p50 was 0.882 / 1.535 / 1.602 seconds
  • Important caveat: evidence is direct and c1 with no controlled decode-rate or routed acceptance; the retained invalid-expectation artifact scored 2/4 because both image rubrics named labels absent from their source assets
  • Decision: retain direct 128K challenger, no-promotion, with the exact 64K recipe as rollback; no route or deployment state changed
  • Canonical evidence: https://fakoli.github.io/anvil-serving/findings/2026-08-17-qwen38-27b-radixark-nvfp4-rtx5090-128k/

X / short post

225 literal characters including the URL as checked on 2026-08-27. Recount immediately before posting.

Local Qwen3.8 27B NVFP4, RTX 5090 128K/c1: 119,675-token retrieval; vision 30/30; media bounds 4/4. Direct only, no promotion. https://fakoli.github.io/anvil-serving/findings/2026-08-17-qwen38-27b-radixark-nvfp4-rtx5090-128k/

Reddit

The title is 71 literal characters. Check the target community's current rules before posting.

Local Qwen3.8 27B NVFP4 on RTX 5090: 128K c1 and media-count boundaries

I tested RadixArk Qwen3.8 27B NVFP4 locally on one RTX 5090 using SGLang,
FP8 E4M3 KV, a 131,072-token window, and concurrency one.

Headline results:

- Long-context retrieval returned the exact marker at 119,675 prompt tokens
  plus 14 completion tokens in 29.811 seconds
- Tool calls passed 20/20
- The established multimodal corpus passed 30/30: image 12/12, mixed 4/4,
  video 14/14
- The corrected media-count corpus passed 4/4: two eight-image and two
  two-video attempts
- Image/mixed/video latency p50 was 0.882/1.535/1.602 seconds

This is direct c1 evidence with no controlled decode-rate or routed-acceptance
result. An earlier artifact is retained at 2/4 because its image rubric named
labels absent from the source assets. This is a bounded local challenger, not
a universal model ranking or a promotion.

Full methodology, failures, and raw artifacts:
https://fakoli.github.io/anvil-serving/findings/2026-08-17-qwen38-27b-radixark-nvfp4-rtx5090-128k/

What matches or differs on your hardware?

Screenshot alt text

Benchmark result card for RadixArk Qwen3.8 27B NVFP4 on one RTX 5090 at 128K context and concurrency one. It reports a 119,675-token retrieval pass, 30 of 30 established multimodal attempts, and 4 of 4 corrected media-count boundary attempts. A caveat notes the evidence is direct c1 and retains an earlier invalid-expectation artifact at 2 of 4.

Claim ledger

Public claim Conditions Evidence
Retrieval pass at 119,675 prompt tokens in 29.811 seconds one direct c1 request; exact marker and stop Finding gates · preflight-needle-119675-tokens.json
Established multimodal 30/30 direct c1 corpus; image 12/12, mixed 4/4, video 14/14 Finding gates · multimodal-c1.json
Corrected media-count boundary 4/4 two eight-image and two two-video attempts at c1 Finding gates · count-boundaries-c1.json
Retained invalid-expectation artifact 2/4 video 2/2; image 0/2 because the rubric named absent source labels Finding failure record · count-boundaries-invalid-expectations.json
Direct-only challenger with no promotion c1; no decode-rate or routed acceptance Finding decision boundary · evidence manifest