Hi all, Some of you have seen my earlier posts about getting NVFP4 working on the Spark, and the tribulations that entailed. When I first got the Spark I felt disheartened by the performance — especially watching Int4 AutoRound models clobber NVFP4 runs (when NVFP4 would even load). Here’s the thi…

Generation config parameters are quite off. Repetition penalty and temp for sure, check Qwen model card. Worse, though, you’ve disabled thinking but left the reasoning parser on. This is probably the root issue. The instruct output is being considered as reasoning.

I could not replicate it. Pulled the stock model from HF and it’s progressing quite nicely. Are you running the tests all concurrently? You may have timeout issues. I’d use Claude or codex as a harness to help identify issues

I ended up getting 5-stars and a score of 90 and even with that I had some timeouts — it scored well enough so that I was convinced there wasn’t a model failure.

Hello there fellow spark owners! I generated a couple of simple quants from the model that was discussed before, for anyone interested: [image] lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled · Hugging Face We’re on a journey to advance and democratize artificial intelligen…

Wonder how the prismaQuant 8bit version compare to the FP8 version…

FYI, this one crashed on me three times last night. I’m now running this FP8 for a couple of hours and so far so good. I couldn’t even benchmark the PQ properly before. This is what the FP8 Benchmark did for me with DFlash with 4 tokens (tried 6 and the effective result was pretty much he same): […

I tried your spark arena recipe, ok it is faster but, quality is worst and TTFT is very bad compare to nvfp4 redhat version Conc. PP128 tok/s TG256 tok/s total TG256 tok/s/req Peak tok/s total Peak tok/s/req TTFR c1 813.97 ± 327.25 53.33 ± 3.47 n/a 65.67 ± 6.60 n/a 205.84 ± 113.21 ms c4 1…

I had no problems with this config on Qwen 3.6 35B because you can activate Thinking via Chat Completion Call later. So i can switch between Thinking and Non-Thinking. Tool-Eval-Bench was fine on 35B. But I removed it as you said and I have these config now and again a bad result on tool-eval-bench: …

Production benchmark: MTP K=3 with FP8 KV + FlashInfer is the deploy-and-forget choice for dense Qwen3.6-27B on DGX Spark I’ve been running rdtand/Qwen3.6-27B-PrismaQuant-5.5bit-vllm on a single DGX Spark (GB10, sm_121a, 128GB unified memory) for an agent workload — 16 concurrent agents, ~30k contex…

Thank you for your post! This reflects my real world observations pretty much exactly. DFlash must run with with full bfloat16 wasting memory and not improving speed at all in real world tasks for me. I am also running this model with MTP3/4 and fp8_e4m3 kv cache :) which is quite usable in the end…

Introducing PrismaQuant

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10

tenari May 4, 2026, 5:25pm 167

Excited to release this!

Topic		Replies	Views
Introducing PrismaScout -- PrismaQuant v2! DGX Spark / GB10	87	5264	June 9, 2026
What's the best speed we can get with Qwen 3.6 27B without quantizing? DGX Spark / GB10	30	14514	June 7, 2026
Qwen3.6-27B is out! DGX Spark / GB10 agentic-ai	284	24301	June 3, 2026
Qwen/Qwen3.5-122B-A10B - Alibaba/Qwen thought about us... :-D DGX Spark / GB10	340	16521	March 24, 2026
Qwen3.5-122B-A10B NVFP4 Quantized for DGX Spark — 234GB → 75GB, Runs on 128GB DGX Spark / GB10 Projects	44	10955	April 9, 2026
RedHatAI/Qwen3.5-122B-A10B-NVFP4 seems to be the best option for a single Spark DGX Spark / GB10 Projects llm	75	6109	May 4, 2026
PSA: State of FP4/NVFP4 Support for DGX Spark in VLLM DGX Spark / GB10	234	12632	May 15, 2026
Qwen/Qwen3.6-35B-A3B (and FP8) has landed DGX Spark / GB10 agentic-ai	308	25826	June 9, 2026
Qwen3.5 27B optimisation thread starting at 30+ t/s TP=1 DGX Spark / GB10 llama , agentic-ai	23	2695	May 11, 2026
Qwen3.5-122B-A10B on single Spark: up to 51 tok/s (v2.1 — patches + quick-start + benchmark) DGX Spark / GB10 cuda , performance , docker , performance-tuning , llm	417	19673	June 9, 2026

Introducing PrismaQuant

Related topics