October 11, 2026 · Models · ~14 min read
DeepSeek V4 Flash vs GLM 5.3 Flash vs Qwen3.8-Flash-Next on a 2× DGX Spark cluster
What's the best model for DGX Spark when you have two of them? We ran the three most talked-about open models of late August 2026 — DeepSeek V4 Flash, GLM 5.3 Flash and Qwen3.8-Flash-Next — on the same 2× DGX Spark cluster and measured them with the same method, under working conditions, with apps running alongside. This isn't a ranking from the makers' benchmarks but our own: DGX Spark tokens per second, how much memory is left, long documents and images, what the license allows — and which silent failures showed up along the way.
Each model has its own page with the full setup and results: DeepSeek V4 Flash, GLM 5.3 Flash and Qwen3.8-Flash-Next. Here we put them side by side.
Short answer: which model for what
Offering to customers · documents
DeepSeek V4 Flash
MIT. Reads long documents fastest (about 2,300 tok/s), most thoroughly tested on documents and as an agent.
Internal use · speed · images
Qwen3.8-Flash-Next
Fastest to start answering, the most free memory, 150.7 tok/s total with 6 conversations. Offering it externally — an agreement with Qwen.
Code and JSON · 1M window
GLM 5.3 Flash
On TensorFold the fastest at code and JSON, with a 1M-token window — but non-commercial only. On vLLM all MIT, and the slowest.
The 2× DGX Spark setup and how we measure
The cluster is two NVIDIA DGX Spark / ASUS Ascent GX10 units with the GB10 chip, 128 GB of unified memory each (the system sees 121.6 GB), linked with a QSFP56 200 Gb/s cable (ConnectX-7, RDMA via NCCL). Each model is split across both units (tensor parallel = 2), one runs at a time, and we switch them from a dashboard in a few minutes. More on the hardware: DGX Spark cluster.
- Median of 6 runs after 3 warm-up requests, never right after startup.
- Temperature 0.6, 400–500-token answers, separately for prose, code and JSON.
- With apps running alongside (a web chat, an API gateway and agents — about 11 GB on the first unit). Recipes online usually measure without them.
- Single runs vary by up to 42% — hence the median, not the best run.
Comparison table
| DeepSeek V4 Flash | GLM 5.3 Flash — vLLM (MIT) | GLM 5.3 Flash — TensorFold | Qwen3.8-Flash-Next | |
|---|---|---|---|---|
| parameters (active) | 284B (13B) | 320B (18B) | 320B (18B) | 125B (6B) + 51B n-grams |
| weights / format | 156 GB, FP8 + DSpark | 184 GiB, NVFP4 | 164 GiB, EXL3 4-bit | 124 GB, NVFP4 |
| prose / code / JSON, tok/s | 44–53 in daily use (series: about 63–65 prose)* | 24.3 / 32.3 / 32.0 | 36.0 / 67.0 / 71.2 | 37.9 / 57.4 / 63.3 |
| several conversations, total | about 150 (6)* | 70.6 (6) | 85.7 (4) | 150.7 (6) |
| time to first token | — | 273–343 ms | 145–184 ms | 142–161 ms |
| input reading | about 2,300 tok/s | about 1,300–1,400 | 1,240–1,690 | — |
| context window | 262K (900K profile) | 262K | 1,048,576 | 262K |
| images | no | yes | yes | yes |
| free after startup (per unit) | about 9–15 GB | about 10 GB | 16–18 GB | 24–26 GB |
| license | MIT | MIT | MIT + non-commercial module | qwen-community (offering → agreement) |
| measured | Aug 18–19 | Sept 23 | Oct 10 | Sept 14 / Oct 8 |
Measured by us, 2× DGX Spark cluster, apps running alongside. * DeepSeek was measured in a different series from the other models (August), so compare it with care.
DGX Spark tokens per second: speed by model
The key observation: the ranking depends on the kind of text. DeepSeek wins on prose, GLM on TensorFold on code and JSON, and Qwen is near the top everywhere while using the least memory.
Generation speed, one conversation
tok/s · measured by us · 2× DGX Spark, apps alongside · ■ DeepSeek V4 Flash (Aug 18 series) ■ GLM vLLM ■ GLM TensorFold ■ Qwen3.8-Flash-Next
prose
code
JSON
DeepSeek: the August 18 series by content type; 44–53 tok/s in daily use, and about 63–65 tok/s on prose in the controlled August 19 series. Other dates: GLM vLLM September 23, GLM TensorFold October 10, Qwen September 14.
Several conversations and time to first token
For a team, total throughput and how quickly answers start matter more than one conversation's speed. Here Qwen and DeepSeek are in a different league from GLM.
Total with several conversations
tok/s in total (number of conversations in brackets) · measured by us
Time to first token
ms, upper end of the range — shorter is better
Long context and memory
For document work, three things count: how long an input the model accepts, how fast it reads it, and how much memory is left for more conversations. Each model wins at something different.
Context window
tokens, in the setups we measured
Input reading
tok/s · longer bar = faster
Free memory after startup
GB per unit · more = more headroom for conversations and apps
In practice: DeepSeek starts commenting on a 100K-token document after about 43 seconds. GLM on TensorFold found information hidden at position 777K, but waited over 10 minutes for its first token. Qwen leaves the most room — which matters, because a GB10 that runs out of memory freezes instead of killing the process.
Images
- Qwen3.8-Flash-Next: 10 images in one message, all described correctly, including small text — 7.9 s without thinking.
- GLM 5.3 Flash: on vLLM, 80/80 images in 167 s, but free memory fell to 2.2 GB (about 40 large photos is safe); it read a Polish invoice in 2.5 s. On TensorFold the limit is 50 images, 50/50 in 16 s.
- DeepSeek V4 Flash in the version we use doesn't accept images.
Licenses — what a company may do
| Model and setup | License | Internal use | Offering to customers |
|---|---|---|---|
| DeepSeek V4 Flash | MIT | ✓ yes | ✓ yes |
| GLM 5.3 Flash — vLLM + NVFP4 + MTP | MIT (throughout) | ✓ yes | ✓ yes |
| GLM 5.3 Flash — TensorFold + DFlash2 | model MIT, DFlash2 module CC-BY-NC-ND-4.0 | non-commercial only | ✗ no |
| Qwen3.8-Flash-Next | qwen-community-1.0 | ✓ yes | a separate license from Qwen first |
A general overview, not legal advice. The terms are covered on each model's page.
Seven lessons from running them
You won't find these in benchmark tables, and they're what cost us the most time.
01
Every model family has its own silent failure
The tool parser (GLM: glm47, not glm), the thinking switch (DeepSeek thinking, Qwen enable_thinking, GLM — none), block-size (GLM 2304). The model looks healthy, but agents go silent, titles come back empty or the content is corrupted.
02
Test accented characters, always
A GLM checkpoint in the ModelOpt format dropped UTF-8 bytes — about 2 broken characters per 10,000. A different checkpoint: 0 errors.
03
A newer engine release can be slower
vLLM 0.31 — Qwen about 30% slower than on the September 14 nightly.
04
The engine matters more than the settings
The same GLM: prose 24 → 36, code 32 → 67 tok/s after switching from vLLM to TensorFold.
05
Measuring is a skill of its own
Up to 42% variation; median of 6 after 3 warm-ups; temperature 0.6 faster than 0 with speculative decoding.
06
GB10's bottleneck is memory, not the link or the GPU
The 200 Gb/s link is 1.66% utilized, the GPU draws 30–35 W. MTU, swappiness, expert parallel — no effect. Speculation works, and its optimum depends on the model (DeepSeek 5 instead of the recipe's 7; GLM and Qwen MTP 3).
07
Online recipes measure without apps alongside
A GB10 that runs out of memory freezes — our GLM runs at gpu-memory-utilization 0.83 instead of the recipes' 0.85. With Qwen, 0.75 gave a bigger pool but unstable speed — we stayed at 0.70.
Which is the best model for 2× DGX Spark
- A company offering the model to customers or partners: DeepSeek V4 Flash or GLM 5.3 Flash on vLLM — both MIT. DeepSeek reads long documents fastest and is the most thoroughly tested on documents and agent work. GLM, according to its maker, is stronger at code and agent work, but on vLLM it's the slowest.
- Internal use, speed and images: Qwen3.8-Flash-Next — fastest to start answering, the most free memory, 150.7 tok/s total with 6 conversations. For external use — an agreement with Qwen.
- Fastest code and JSON, a 1M window: GLM 5.3 Flash on TensorFold — but non-commercial only (the DFlash2 module).
Our cluster runs one large model at a time, and all apps call one shared model name — so switching models takes a few minutes, without rewriting the apps. If you're still choosing hardware, see also DGX Station vs a DGX Spark cluster.
Frequently asked questions
What is the fastest model on DGX Spark?
On our 2× DGX Spark cluster it depends on what you measure. Qwen3.8-Flash-Next has the shortest wait for the first token (142–161 ms), and together with DeepSeek V4 Flash it delivers the most in total with six conversations (about 150 tok/s). GLM 5.3 Flash on the TensorFold engine writes code and JSON fastest (67.0 and 71.2 tok/s), and DeepSeek V4 Flash reads long documents fastest (about 2,300 tok/s).
What is the best model for coding on DGX Spark?
GLM 5.3 Flash on TensorFold writes code fastest (67.0 tok/s), but that setup is non-commercial only. For a company: DeepSeek V4 Flash (56.3 tok/s on code, tested as a coding agent, MIT license) or, for internal use, Qwen3.8-Flash-Next (57.4 tok/s).
Do these models need two DGX Sparks?
In the versions we measured, yes — all three run on two 128 GB units. Only more heavily compressed variants fit on one unit according to their sources: DeepSeek's 3-bit GGUF (about 103 GB), GLM's GGUF Q2 or EXL3 2.05 bpw (about 109 and 85 GB), and Qwen per the recipe author on one unit with the n-gram table read from disk (43.9 tok/s). We haven't tested those variants.
Which model can be offered to customers?
DeepSeek V4 Flash and GLM 5.3 Flash on vLLM — both MIT throughout. Qwen3.8-Flash-Next requires a separate license from Qwen before offering the model to third parties, and the GLM setup on TensorFold uses the DFlash2 module under a non-commercial license.
Sources
- Measurements and tests: our own, ailone's 2× DGX Spark / ASUS GX10 cluster — DeepSeek V4 Flash August 18–19, GLM 5.3 Flash September 23–24 (vLLM) and October 10 (TensorFold), Qwen3.8-Flash-Next September 14–15 and October 8, 2026.
- Full setups, results and sources for each model (model cards, checkpoints, community recipes, vLLM issues, licenses): DeepSeek V4 Flash, GLM 5.3 Flash, Qwen3.8-Flash-Next.
- Community recipes for 2× DGX Spark we built on: MiaAI-Lab (DeepSeek), tonyd2wild (GLM), MiaAI-Lab (GLM, TensorFold), tonyd2wild (Qwen).
As of October 11, 2026. Engines and versions change every week — the numbers apply to the setups described on the model pages. Measurements are medians from one series, with apps running alongside.