Qwen3.8 27B RadixArk NVFP4 on RTX 5090

Date: 2026-08-17 Decision: challenger, no-promotion Evidence: functional, capacity, bounded deterministic quality 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 fit one RTX 5090 at 65,536 context and concurrency one, then passed the retained long-context, tool-call, and direct multimodal gates without changing any route or promotion state.

Setup	Qualified value
Model	RadixArk/Qwen3.8-27B-NVFP4@554ebba9b5f1b79dc11246341960360e6ef05ef4, served as qwen38-27b-radixark-nvfp4-sglang-rtx5090-64k-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 64K multimodal recipe at 85b21147
Measurement path	direct online managed endpoint; retained artifacts do not classify individual requests as cold or warm
Contract	65,536 context, c1, 512-token multimodal output cap, up to four images per corpus request, one video per corpus case, thinking disabled
Evidence	functional, capacity, bounded deterministic quality; retained qualification complete
Decision	challenger, no-promotion; no route, deployment, or live serving state changed

Headline measurement	Local result	Conditions
Long-context retrieval and tools	pass; 20/20 tools	approximately 60,000 prompt tokens, exact marker, stop; c1
Deterministic multimodal corpus	30/30	image 12/12, mixed 4/4, video 14/14; c1
Direct modality latency p50	0.886 / 1.548 / 1.648 s	image / mixed / video across the 30-attempt corpus
Untested boundary	no result	controlled decode rate, c2+, routed acceptance, or action loop

Why it matters: This is a reproducible single-card local perception, long-context, and tool-formatting challenger with an exact managed rollback surface.

Important caveat: The retained run is direct and concurrency one; it does not establish controlled decode throughput, routed media admission, broad GUI grounding, or end-to-end computer action.

Evidence manifest · Publication summary

Outcome

The digest-pinned RadixArk Qwen3.8 27B ModelOpt NVFP4 checkpoint fits a single RTX 5090 and is the preferred locally proven 5090 computer-use perception candidate. It passed text, structured output, a 60K retrieval needle, 20/20 tool calls, direct image/OCR/video preflight, and the complete deterministic multimodal corpus at 30/30. The corpus includes temporal order, state change, event localization, video OCR, 120-second continuity, multi-image comparison, mixed video-plus-image requests, and two supplementary real-world clips.

This result does not promote or route the model. It establishes a reproducible single-card challenger and closes the need to download the weaker llama.cpp frame-sampling fallback for this qualification round.

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.
• FlashInfer build: 0.6.18.dev20260807, source revision 906181e3f4cf4bcc81835fb480db4011bbd80b62.
• Recipe: configs/qwen38-27b-radixark-nvfp4-sglang-rtx5090-64k-mm-recipe.toml.
• Weight format: ModelOpt mixed precision, NVFP4 dense MLP paths, FP8 projections/attention paths, and BF16 vision/MTP tensors.
• KV: FP8 E4M3. Context: 65,536. Concurrency: one. MTP: disabled for the stability baseline. Multimodal feature transport: CPU.
• Cache snapshot: 21,945,295,265 logical bytes, zero incomplete files, zero broken links, and only the requested revision.

The managed load reported 30.08 GB available before weights, 20.14 GB used by the model, 9.94 GB available after weights, and 3.73 GB available after KV, eager buffers, and the decode CUDA graph. The engine allocated 167,789 KV tokens even though the admission profile is capped at 65,536.

Gates and measurements

Gate	Result
Coding and structured JSON	pass
Retrieval needle	pass at approximately 60,000 prompt tokens; exact marker returned with stop
Tool calls	20/20 valid calls with tool_calls finishes
Thinking contract	explicitly disabled; no reasoning-channel output observed
Direct image understanding	pass with required text and stop
Direct verbatim OCR	pass with ANVIL READY, 42917, and 42
Direct MP4 video	pass with ordered red then green
Complete multimodal corpus	30/30; image 12/12, mixed 4/4, video 14/14

The complete corpus ran at concurrency one with two repetitions per case, a 512-token output cap, thinking disabled, and only stop accepted. Aggregate latencies were 0.886 seconds image p50 / 1.977 p95, 1.548 seconds mixed p50 / 2.232 p95, and 1.648 seconds video p50. Video p95 was 142.051 seconds because the two supplementary real-world clips generated long descriptive answers; the five deterministic synthetic video cases each remained at or below 2.097 seconds p95.

Raw, sanitized evidence:

• functional, 60K, and tools preflight
• direct image/OCR/video preflight
• 30-request deterministic multimodal corpus
• source registry

Failures and caveats

1. The first generic preflight used its default approximately 128K needle and correctly received HTTP 400 from the deliberately 65,536-token profile. The declared-profile reruns passed at 48K and 60K. This is retained as a harness/configuration mismatch, not a model failure.
2. The first direct multimodal preflight used a 256-token visible-output cap. The required image content was present, but the descriptive answer ended with length; the bounded 512-token rerun ended with stop and passed.
3. SGLang warned that FP8 KV scaling factors were absent and defaulted to 1.0. The deterministic gates passed, but equivalence to unquantized KV is not established.
4. The generic eval benchmark evidence show command does not recognize the multimodal evidence schema. The artifact was inspected directly and the product gap is recorded in a repository ticket.
5. No controlled text decode-rate benchmark, concurrency above one, routed acceptance, broad GUI-grounding benchmark, or computer-action loop was run. The result proves perception, temporal video handling, long-context retrieval, and tool-call formatting—not end-to-end autonomous computer use.

LFM2.5-VL and GGUF fallback comparison

LiquidAI/LFM2.5-VL-1.6B@919fde3d022e3f90a4716006f993938ee8c2eb97 is a strong small-model companion: its official card recommends it for general vision, OCR, and document comprehension, and the official WebGPU demo processes live video frames locally. Its 1.6B size makes it attractive as an always-on screen watcher, frame captioner, OCR prefilter, or structured event extractor. It was not locally benchmarked in this run.

It is not the primary recommendation for this goal. Its documented context is 32,768 tokens, its published video demo processes frames, and the reviewed official material does not establish native temporal video-file reasoning, GUI grounding, or computer-action reliability. Qwen3.8 NVFP4 directly accepted MP4 input here, passed temporal and continuity assertions, and retained a 64K served window.

The prepared fallback was unsloth/Qwen3.8-27B-GGUF at f1bfb127c64f7072bdd2cad55f258b9c8b2910fe with Q4/UD-Q4 weights and an F16/BF16 multimodal projector. Current llama.cpp multimodal documentation supports image and experimental audio through libmtmd, not native video; that path would require explicit frame extraction and multi-image prompting. Because the native SGLang NVFP4 candidate passed, the fallback was not pulled or locally measured.

Decision boundary

Retain the exact recipe as the preferred RTX 5090 multimodal challenger and keep the checkpoint cached. Do not route or promote it without a separate human gate, routed admission limits, client acceptance, GUI-grounding/action evaluation, and a decision on the FP8-KV scale warning. LFM2.5-VL remains a promising lightweight auxiliary model, not a replacement for this native video-capable reasoning lane.