Qwen 3.8 27B on One GX10: The Receipts

I cheered Qwen 3.8 27B before the weights existed. Then I loaded NVFP4 on the ASUS Ascent GX10 and timed it. This is that week, with the threads and the comments attached.

The bet

The Aug 10 cheer — Qwen 3.8 27b, thirty times, then “We believe in you!!!!”

That post did ~538K impressions. I later quoted it and wrote, “I wish I could still write these bangers like I used to.” @decapostos asked for the secret. I told him I’m not sure it’s repeatable.

QwenDevs answered the cheer in-thread: “Got you covered bro, 27B is landing this week.” Weights did. Official repo is Qwen/Qwen3.8-27B, Apache 2.0, BF16 about 55.6 GB, 262K native context, vision included. The file I actually run is NVFP4, not the 55 GB safetensors.

The first-look note is still up: Qwen 3.8 27B Locally: We Believe in You. That was the load plan. The run started after weights landed Aug 13–14.

Hardware is the same ASUS Ascent GX10 I unboxed Aug 8: The ASUS Ascent GX10 Showed Up — NVIDIA GB10, 128 GB unified. One box. Not a cluster.

What I actually ran

Quant: Qwen3.8-27B-NVFP4.

Stacks I named in public:

1. Unsloth NVFP4 on vLLM + MTP (:8000)
2. RadixArk ModelOpt NVFP4 on SGLang + EAGLE (:8888)
3. sparkrun vLLM 0.27 + MTP-3 after @jaita pointed at styles01/sparkrun-recipes and the drowzeys GB10 image (:8078)

Baseline → MTP → DSpark sweep, thinking off, temp 0, wall-clock tok/s:

Seat	short	med	long
Baseline Unsloth NVFP4	7.3	11.0	11.2
MTP Unsloth NVFP4	5.6 (0.77×)	19.0 (1.73×)	25.5 (2.28×)
DSpark Unsloth NVFP4	0.8	10.2	8.3
DSpark ModelOpt NVFP4	5.3	11.5	12.2
FP8 + RadixArk DSpark	4.0	7.87	8.52

DSpark accept rate 3.4–8.1%, mean accept length 1.24–1.57. MTP accept 60–100% on longer sequences, mean accept length 2.2–3.0. MTP won.

MTP depth sweep n=1–4, thinking off. Daily winner n=3: med 19.7 / long 20.0 tok/s, math 11.5, accept ~56%, mean accept length 2.67. n=4 added about 7% on long (21.5) and lost short/math. Accept fell to ~51%. Default MTP_N_SPEC=3.

@jurlycat said n=4’s long-context gain probably isn’t worth the acceptance hit for mixed workloads.

@neojohnson113 asked if 27B could sit next to DeepSeek V4 Flash on one Spark. I said I don’t think it can fit. Flash stays the Moon Trail coding loop: Hermes + DeepSeek V4 Flash on the GX10.

The speed-leader thread

We have a new speed leader.

Matched single-stream, same host, one heavy stack at a time, thinking off, smoke 256 / test1 4096.

	vLLM + MTP · Unsloth	SGLang + EAGLE · RadixArk
Smoke decode	25.89 tok/s	32.19 tok/s (+24%)
Test1 decode	26.09 tok/s	31.19 tok/s (+20%)
Smoke TTFT	1.46s	0.35s
Test1 TTFT	0.47s	0.27s (~1.8×)

SGLang won. Recipe claimed ~33–35 tok/s single. We landed ~31–32. Tagged @MiaAI_lab and @sgl_project.

Comments from that thread:

• @MiaAI_lab: “Yeah sglang is faster on this model vs vllm.”
• @jaita: “I got 32 on vLLM.”
• @PhilShteuck: bake-offs have a shelf life of about nine days.
• @jtregunna on 2× RTX A6000 (not a Spark): vLLM BF16, 8-bit KV, MTP 2, up to 44 tok/s high / 30 low, ~980 tok/s prefill; SGLang ~12% slower on decode. Caveat: different card.
• @unseenmars_ asked about DSpark.
• @DJLougen pointed at vanilla SGLang + DSpark, quoting sgl_project (206.1 tok/s 5090, 38.28 DGX Spark).
• @alexellisuk wanted prefill and a longer sweep.
• @RedmixGeco on 4.2× smoke TTFT. I said Asus Ascent GX10.
• @JakeSmGaming QT: “SGLang keeps winning these. TTFT from 1.46s to 0.35s is wild.”
• @vernons: Lightning ~100 tok/s vs Qwen 3.8 NVFP4 ~10 tok/s, MoE-vs-dense confound.

The long-HTML bake-off

Follow-up bake-off.

PULSE-16 locked prompt SHA 4cae231f…, thinking off, max_tokens 64k, long HTML.

Seat	tok/s	TTFT	tokens	gen
A prior vLLM+MTP (Unsloth)	21.78	0.84s	23,606	1083.6s
B SGLang+EAGLE (RadixArk)	24.62	0.65s	19,526	793.2s
C sparkrun vLLM 0.27 MTP-3	21.52	0.67s	25,665	1192.7s

B still fastest (+13% vs A). C is A-class and gate-passed. External “~32 on vLLM” did not show up as ~32 on long HTML. C smoke ~21. Short TG claims ≠ long HTML TPS.

@jaita is in that thread.

What I used 27B for

• Breakout
• Racer
• Nemotron same prompt
• Laguna same prompt
• Cheer-thread reply game
• Speed test reply
• Same-prompt clip dump
• @jaita asked about Hermes quirks. That question is still open. I am not inventing an answer.

Pirate-ship caveat

I ran the same “build a pirate ship game” brief through 6 models. Content, not science. His own bullets:

• Same prompt ≠ same product. Each model invents a different game, so you’re ranking vibes, not capability on a fixed task.
• The footage is scripted autopilot (orbit + force-click), not a human playing. Smooth video ≠ fun.
• Local vs cloud confounds stack, quant, thinking flags, max tokens, and parent polish.
• One GPU at a time → sequential, not fair head-to-head latency.
• Pretty six-panels with burn-in labels hide broken meshes, bare identifiers, and “works on my capture harness.”
• Freeform Test-1 energy: brands and UIs diverge; the bench quietly becomes a design contest.

What would be a real bench: locked identity + identical prompt bytes + completeness gate + honest autopilot vs organic labels + stack IDs only. Until then this is content, not science.

35B-A3B

This will be next, quoting AstraiaAI. Next-weight note, not a 27B receipt.

How I would load it this week

1. One GX10. Do not pair 27B NVFP4 with Flash as two residents.
2. Start Unsloth NVFP4 on vLLM, MTP n=3, thinking off.
3. If first-token latency matters, time RadixArk SGLang+EAGLE on your prompt.
4. If someone quotes ~32 tok/s, ask smoke-256 vs 20k-token HTML.
5. Native NVFP4 on Blackwell. Ada/Ampere still Q4_K_M. See the VRAM requirements guide and the 2026 model tracker.
6. Flash for the long coding loop. 27B for chat and light agents.

If you timed the same quant on a Spark, a 4090, or something else, send the stack and the prompt class — X (@wikiwayne).