⚡ Update: v2 (post #71) achieves 51 tok/s. v2.1 (post #104) adds a quick-start script. See those posts for the latest setup. Been chasing every last token/second out of Qwen3.5-122B-A10B on a single DGX Spark for the past few weeks. Not sure if anyone else is still optimizing this model on Spark, b…

When it comes to quantization in general, I agree. For example, MXFP4 and others are great. I am talking just about int4-AutoRound. Very fast, great for DGX Spark. But for my use cases, it makes Qwen3.5-122b perform much worse quality wise. Unfortunatly bad enough to skip it. My first unstructure…

Thanks a lot!! You inspired me testing it. And I agree, incredible throughput. Quality seems much better than int4-Autoround for the 122b variant…even though I don’t want to believe this

There was definitely some kind of bug in vLLM within the last week which caused poor output from int4-Auroround quants. It seems to now have been fixed. Consider trying int4 again. This hybrid repo’s pins do not seem to have the problem.

Looking forward to try your solution.

Would you share your MTP settings and startup script?

Your finding matches my real-world experience exactly. When using 122B INT4 for coding tasks, I noticed it would sometimes enter very long thinking loops and occasionally get stuck without producing any output — especially on complex multi-step problems. 35B Native FP8 just… gets it done. Faster th…

[image] chrm: I wonder, can we add the great 51 tok/s solution concept to your docker community image or it this solution already working as tested by myself (although I have not compared the mtp accuracy, but both are working for programming)? I don’t think any modifications for the docker b…

In brief, this thread pins vllm=0.19.0 because it depends on two different patches. vLLM’s architecture has changed since release and the patches - definitely the one to inc.py are not forward compatible right now. One of them is applied at build time for the hybrid FP8+int4 model. It is required f…

I think this patch is needed in your docker build to get extra tps boost: DGX_Spark_Qwen3.5-122B-A10B-AR-INT4/patches/03-int8-lm-head/patch_int8_lmhead.py at master · albond/DGX_Spark_Qwen3.5-122B-A10B-AR-INT4 · GitHub

Qwen3.5-122B-A10B on single Spark: up to 51 tok/s (v2.1 — patches + quick-start + benchmark)

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

joshua.dale.warner April 12, 2026, 9:28pm 241

@bernardlbmi3 @stefan132 @paxren2020

I have made a separate thread to discuss work around and optimizations of Qwen3.5-35B-A3B.

Let’s continue at Qwen3.5-35B-A3B optimizations on single Spark

Topic		Replies	Views
Qwen/Qwen3.5-122B-A10B - Alibaba/Qwen thought about us... :-D DGX Spark / GB10	340	16025	March 24, 2026
Qwen/Qwen3.6-35B-A3B (and FP8) has landed DGX Spark / GB10 agentic-ai	251	21541	May 24, 2026
Qwen3.5-122B-A10B NVFP4 Quantized for DGX Spark — 234GB → 75GB, Runs on 128GB DGX Spark / GB10 Projects	44	10349	April 9, 2026
Does Qwen3.5-35B-A3B on GB10 leave a lot of performance on the table? DGX Spark / GB10 agentic-ai	40	5547	March 16, 2026
Qwen3.5-35B-A3B optimizations on single Spark DGX Spark / GB10 Projects	48	2775	May 22, 2026
Qwen3.5-122B-A10B on single Spark: 15 → 21.5 tok/s with hybrid GPTQ-INT4 + FP8 dense layers (https://github.com/rmstxrx/vllm-hybrid-quant) DGX Spark / GB10 cuda	9	727	March 20, 2026
Qwen3.5-397B-A17B run in dual spark! but I have a concern DGX Spark / GB10	235	8349	May 23, 2026
HOW-TO: Run Qwen3-Coder-Next on Spark DGX Spark / GB10 llama	92	9567	March 24, 2026
RedHatAI/Qwen3.5-122B-A10B-NVFP4 seems to be the best option for a single Spark DGX Spark / GB10 Projects llm	75	5775	May 4, 2026
Qwen3.5-397B-A17B + DGX Spark (duo) DGX Spark / GB10 Projects	56	5321	April 13, 2026

Qwen3.5-122B-A10B on single Spark: up to 51 tok/s (v2.1 — patches + quick-start + benchmark)

Related topics