Qwen3.8 Flash Next and GLM 5.3 Flash on DGX Spark: first speeds

By 6 September owners on NVIDIA's forum had Qwen3.8 Flash Next running on a single DGX Spark at 37 to 44 tokens a second, and one post had GLM 5.3 Flash on one box at 64. The 64 came from a build squeezed to about 2 bits per weight, and I would want to see the quality numbers before copying it.

By Alex · 6 September 2026 · 2 minute read

The video: RTX Spark Ships in October, and NVIDIA Buys Hugging Face — This story starts 3 minutes 18 seconds into my video of 6 September 2026, RTX Spark Ships in October, and NVIDIA Buys Hugging Face

By 6 September, several owners on NVIDIA's developer forum had Qwen3.8 Flash Next running on a single DGX Spark. Their figures were 37 to 44 tokens a second, depending on the recipe and how much context they left room for.

The models

Flash Next came out on Wednesday 26 August, the same day as GLM 5.3 Flash, the model that went onto my four DGX Sparks on 28 August. Two days later came the giants, Tencent's Hy4 preview and the full GLM 5.3. Tencent's model card puts Hy4 preview at 770 billion parameters, with 49 billion active for each token, and the full GLM 5.3 is a similar size. Both are open weights and neither fits in one box.

On Artificial Analysis's Intelligence Index, as shown on 5 September, the two Flash models scored 46 each and the full GLM 5.3 scored 49 at its max setting. Hy4 preview was not listed yet.

The first speeds

Model	Machines	Speed	Conditions	Whose figure
Qwen3.8 Flash Next	1 DGX Spark	43.81 tok/s	median, single request, 4-bit build, 8K context	an owner
Qwen3.8 Flash Next	1 DGX Spark	about 43 tok/s	coding	an owner
Qwen3.8 Flash Next	1 DGX Spark	about 37 tok/s	prose, 1M context	an owner
GLM 5.3 Flash	1 DGX Spark	64 tok/s	structured output, 2.05 bits per weight	an owner
GLM 5.3 Flash	1 DGX Spark	25 tok/s	ordinary prose, the same build	the same owner
GLM 5.3 Flash	4 DGX Sparks	46.3 tok/s	single user, 4-bit NVFP4 build	measured by me
The full GLM 5.3	4 DGX Sparks	12.12 tok/s	one user, 200K context	an owner

Owners' figures are from NVIDIA's developer forum, 29 August to 5 September 2026. My own is from 28 August 2026.

Separately, llama.cpp version 0.4.0 came out on 4 September, and its release notes say it adds the first support for Flash Next.

GLM 5.3 Flash on one box

One post, from 3 September, had GLM 5.3 Flash running at 64 tokens a second on a single Spark, but there were two catches. That build squeezes the model down to 2.05 bits per weight, and the 64 is for structured output. The same post gives 25 tokens a second for ordinary prose.

As a comparison, my own four DGX Spark cluster was running the same model as a 4-bit NVFP4 build, spread across all four boxes. I measured it at 46.3 tokens a second for a single user on 28 August.

The two giants

By my own maths, Hy4 preview and the full GLM 5.3 would each still be about 390GB at 4 bits, against 128GB in one Spark and 512GB in four. One owner reported the full GLM 5.3 on four Sparks at 12.12 tokens a second for one user and 46.03 in total across six, and said no formal evaluation against the original had been run. When I searched the forum on 5 September I found two passing mentions of Hy4, and no recipe or numbers.

What I would do

A 2-bit build of GLM 5.3 Flash on one box is a real achievement, but I would want to see the quality numbers before copying it.

An owner's Flash Next speeds on one DGX Spark, 2 Septemberforums.developer.nvidia.com
A second thread of Flash Next speeds on one DGX Spark, 30 Augustforums.developer.nvidia.com
A third thread of Flash Next speeds on one DGX Spark, 5 Septemberforums.developer.nvidia.com
The post with GLM 5.3 Flash on one DGX Spark, 3 Septemberforums.developer.nvidia.com
An owner's post with the full GLM 5.3 on four DGX Sparks, 29 Augustforums.developer.nvidia.com
Artificial Analysis, open-weights leaderboardartificialanalysis.ai
The llama.cpp 0.4.0 release notes on GitHubgithub.com
Qwen's model card for Qwen3.8 Flash Nexthuggingface.co

The numbers in this story

Figure	What it is	From	Kind
46, 46 and 49	The two Flash models and the full GLM 5.3, as shown on 5 September	artificialanalysis.ai	Artificial Analysis's score
43.81, about 43 and about 37 tok/s	Flash Next on one DGX Spark, from three forum threads	forums.developer.nvidia.com	an owner's report
64 and 25	GLM 5.3 Flash on one DGX Spark, structured output and prose	forums.developer.nvidia.com	an owner's report
46.3 tok/s	GLM 5.3 Flash on my four DGX Sparks, for a single user	my own test	measured by me
12.12 and 46.03	The full GLM 5.3 on four DGX Sparks, for one user and across six	forums.developer.nvidia.com	an owner's report
770B and 49B	Hy4 preview, parameters in total and active for each token	huggingface.co	Tencent's model card

My Sparks would run a little faster than these figures. I cap their graphics chips at 2,200MHz, against about 2,400 uncapped, to cut their power use. When I tested the cap on one box in August, writing speed stayed within 1 per cent, reading was 2 to 4 per cent slower, and the chip's peak power fell from 55.5W to 43.3W.