Skip to content

Qwen3.8 27B RadixArk NVFP4 128K qualification on RTX 5090

Date: 2026-08-17 Decision: retain direct 128K challenger; no-promotion Evidence: functional, bounded deterministic quality, multimodal boundary Measured hardware: one NVIDIA GeForce RTX 5090, 32,607 MiB, sm_120 Topology: isolated Windows 11 / Docker Desktop / WSL2 qualification lane; no RTX PRO 6000 was measured, protected, or involved

Result card

RadixArk Qwen3.8 27B NVFP4 retained its direct RTX 5090 challenger status at 131,072 context after passing a 119,675-token retrieval request, the 30-case multimodal corpus, and the eight-image/two-video boundary corpus at c1.

Setup Qualified value
Model RadixArk/Qwen3.8-27B-NVFP4@554ebba9b5f1b79dc11246341960360e6ef05ef4, served as qwen38-27b-radixark-nvfp4-sglang-rtx5090-128k-mm
Hardware one NVIDIA GeForce RTX 5090, 32,607 MiB, sm_120; isolated Windows 11 / Docker Desktop / WSL2 lane
Runtime SGLang c4271c3fe1262fc2adbd162c33b25de5255251c5; image sha256:506525a5907ea22c9d445afb7c03603959b912de034d86915cf17da814f1a124; ModelOpt NVFP4/FP8 weights, FP8 E4M3 KV, MTP 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
Contract 131,072 context, 131,066 maximum input, c1, 512-token multimodal output cap, eight images or two videos, thinking disabled
Evidence functional, bounded deterministic quality, multimodal boundary; retained qualification complete
Decision retain direct 128K challenger, no-promotion; no route or deployment state changed; retain 64K rollback
Headline measurement Local result Conditions
Long-context retrieval pass in 29.811 s 119,675 prompt plus 14 completion tokens, exact marker, stop; c1
Established multimodal corpus 30/30 image 12/12, mixed 4/4, video 14/14; c1
Media-count boundary corpus 4/4 two eight-image and two two-video attempts; c1
Direct modality latency p50 0.882 / 1.535 / 1.602 s image / mixed / video across the 30-attempt corpus
Retained invalid-expectation artifact 2/4 both video attempts passed; both image rubrics named absent source labels

Why it matters: The exact managed 128K recipe preserves direct local long-context and multimodal coverage while recording a concrete media-count boundary and an exact 64K rollback.

Important caveat: This remains direct, concurrency-one evidence with no controlled decode-rate benchmark or routed acceptance; the first image-boundary artifact failed because its rubric required labels absent from the source images and is retained alongside the corrected 4/4 run.

Evidence manifest ยท Publication summary

Outcome

The separate 128K managed recipe passed its retention gates and replaced the 64K candidate on the direct qualification endpoint. The server advertises a 131,072-token context window, a 131,066-token maximum input, a 167,789-token KV pool, and explicit per-request ceilings of eight images and two videos. No router alias or deployment promotion changed; the 64K recipe remains the exact rollback.

The model returned the retrieval marker with 119,675 actual prompt tokens, completed tools 20/20, passed direct image/OCR/video preflight, passed the complete established multimodal corpus 30/30, and passed the new count-boundary corpus 4/4: two eight-image attempts and two two-video attempts.

Immutable identity and recipe

  • Model: RadixArk/Qwen3.8-27B-NVFP4 at 554ebba9b5f1b79dc11246341960360e6ef05ef4.
  • Runtime: lmsysorg/sglang:qwen38-27b at image digest sha256:506525a5907ea22c9d445afb7c03603959b912de034d86915cf17da814f1a124.
  • SGLang source revision from the image label: c4271c3fe1262fc2adbd162c33b25de5255251c5.
  • Managed recipe: configs/qwen38-27b-radixark-nvfp4-sglang-rtx5090-128k-mm-recipe.toml.
  • Served name: qwen38-27b-radixark-nvfp4-sglang-rtx5090-128k-mm.
  • ModelOpt NVFP4/FP8 mixed weights; FP8 E4M3 KV; TP=1; concurrency one; FlashInfer attention; 2,048-token chunks; CPU multimodal feature transport; radix cache disabled; one Mamba/GDN state slot; MTP disabled; thinking disabled.
  • Exact cached snapshot: 21,945,295,265 logical bytes, with zero incomplete files or broken links.

Startup reported 30.08 GB available before weights and the same 20.14 GB model footprint as the 64K baseline. After startup, the host reported 3,928 MiB free. The engine retained the measured 167,789-token KV pool while raising request admission from 65,536 to 131,072 tokens.

Gates and measurements

Gate Result
Coding and structured JSON pass
Long-context retrieval pass; 119,675 prompt tokens plus 14 completion tokens, exact marker, stop, 29.811 s
Tool calls 20/20 valid, tool_calls finishes
Thinking contract explicitly disabled; reasoning output forbidden and absent
Direct image / OCR / MP4 all pass with required labels and stop
Established multimodal corpus 30/30: image 12/12, mixed 4/4, video 14/14
Eight-image boundary 2/2; ordered evidence from all eight positions
Two-video boundary 2/2; both clips' state transitions returned in supplied order

The established corpus used concurrency one, a 512-token output cap, four-image and one-video evidence ceilings, thinking disabled, and only stop accepted. Image latency was 0.882 seconds p50 / 1.973 p95; mixed was 1.535 / 2.222; video was 1.602 / 134.440. The long video p95 is dominated by two descriptive real-world clips. The boundary corpus used recorded ceilings of eight images and two videos: eight-image p50/p95 was 3.803/3.927 seconds and two-video was 3.119/3.138 seconds.

Raw, sanitized evidence:

Failure record and caveats

  1. The first count-boundary artifact passed the two-video attempts but failed both eight-image assertions. Inspection proved the corpus asked for labels absent from its hash-pinned source images (Q4, submit, and a blue circle). The model correctly described the actual Q1-Q3 chart, status/GPU panel, and three blue boxes. The corrected corpus uses the assets' canonical labels and a repeated ordered code to cover the first and eighth images; it passed 4/4 under a new corpus hash. Both artifacts are retained.
  2. The benchmark harness previously hard-coded one video per case. The managed surface now accepts a fail-closed --max-videos-per-request value from one through sixteen and records it in plans and artifacts. This is evidence admission only; it does not modify the engine or router.
  3. The preflight generator's requested --needle-ctx 120000 produced 99,049 actual prompt tokens. A separate larger bounded request produced 119,675 actual prompt tokens and is the basis for the long-context claim.
  4. SGLang still warns that absent FP8 KV scaling factors default to 1.0. Deterministic gates passed, but equivalence to unquantized KV is unproven.
  5. The result remains concurrency-one, direct-endpoint evidence. It does not establish routed media admission, controlled decode rate, broad GUI grounding, or a closed-loop computer-action policy.

Decision boundary

Retain the exact 128K recipe as the active direct RTX 5090 vision/video qualification candidate and retain the 64K recipe as rollback. The proven request boundary is eight images or two videos, not an unlimited media count. Do not route or promote without a separate human gate, router admission limits, client acceptance, GUI-grounding/action evaluation, and a decision on the FP8 KV scaling warning.