Best models for DGX Spark, October 2026 edition
On one DGX Spark, Qwen3.6-35B-A3B in NVFP4 with an MTP draft is the fastest current pick, reported at 86.3 tok/s single-stream at a 4k context fill. On two Sparks, DeepSeek-V4-Flash runs its official FP8 weights at about 44 tok/s. The tables below give one pick per use case, with the quantization, engine, source and date behind each number. Every figure is reported by a third party unless it says GPUwerk-measured. Sources read 24 October 2026.
Verdict: 1 Spark
| Use case | Model | Quant | Engine | Single-stream tok/s | Source | Date |
|---|---|---|---|---|---|---|
| Fastest chat and agents | Qwen3.6-35B-A3B | NVFP4 + MTP (3-token draft) | vLLM, eugr build | 86.3 at 4k fill (advantage) | SparkBench | 2026-08-04 |
| Coding, GPUwerk-measured | Qwen3-Coder-30B-A3B | AWQ 4-bit | vLLM 0.27.1 | 80.9 | GPUwerk benchmarks | 2026-09-02 |
| Small and efficient | Nemotron 3 Nano 30B-A3B | NVFP4 (W4A16) | vLLM 0.20.1, patched | 74.75 | ai-muninn | 2026-05-01 |
| General and vision | Gemma 4 26B-A4B | NVFP4 + MTP | vLLM 0.26.1 nightly | 54.9 | NVIDIA forum | 2026-08-10 |
| Dense 27B | Qwen3.8-27B | NVFP4 + DFlash2 (8-token draft) | SGLang | 38.1 at 4k fill | SparkBench | 2026-09-10 |
| Dense 27B, older release | Qwen3.6-27B | NVFP4 | not stated | 33 | noze, citing MiaAI-Lab | 2026-08-06 |
| Batch and agent fleets | gpt-oss-120b | MXFP4 | vLLM 26.03 container | 33.5 (862.8 aggregate at 256) | Dendro Logic | 2026-04-22 |
| Largest model, cut to fit | DeepSeek-V4-Flash | 2-bit | ds4 fork for GB10 | ~28 | NVIDIA forum | 2026-08-01 |
| GLM on one box | GLM-5.3-Flash (320B-A18B) | UD-Q2_K_XL, 108.72 GB | llama.cpp build 10677 | 15.4 prose / 23.6 code | ai-muninn | 2026-08-30 |
SparkBench figures are decode after a fixed context fill. Qwen3.6-35B-A3B drops to 78.8 tok/s at 50k and 31.5 at 100k, and Qwen3.8-27B to 19.1 at 50k. Other one-Spark reports for Qwen3.6-35B-A3B are 81 tok/s (NVFP4, noze, citing MiaAI-Lab) and ~106 tok/s with an Unsloth speculative-decoding recipe (openclawdc). Nemotron 3 Nano is also reported at ~67 tok/s on an earlier community vLLM image (NVIDIA forum). DeepSeek-V4-Flash on one Spark is reported from ~13 tok/s (patched vLLM fork, prose) to ~55 typical (reduced NVFP4, openclawdc). Speculative decoding drafts several tokens per weight read, so its gain depends on how predictable your text is.
Verdict: 2 Sparks
| Use case | Model | Quant | Engine | Single-stream tok/s | Source | Date |
|---|---|---|---|---|---|---|
| Frontier MoE, official weights | DeepSeek-V4-Flash (284B, 13B active) | official FP8 + MTP | vLLM (jasl fork), TP=2 | ~44 (advantage) | NVIDIA forum | 2026-05-16 |
| Very long context | DeepSeek-V4-Flash | official checkpoint, FP8 KV, 1M context | vLLM 0.21.1, TP=2 | 45.7 empty, 30.4 at 980K | NVIDIA forum | 2026-06-04 |
| GLM-5.x | GLM-5.2 | 2-bit experts, 96K context | TP=2, engine not named | ~25.8 | openclawdc | 2026-08-05 |
| 235B MoE at 4-bit | Qwen3-235B-A22B | UD-Q4_K_XL, 134 GB | llama.cpp RPC | ~12.5 (about 1.8 on one Spark) | NVIDIA forum | 2025-12-17 |
The vLLM recipe for DeepSeek-V4-Flash (updated 29 September 2026) uses two GB10 nodes with TP=2 and says the checkpoint does not fit one Spark’s 128 GB. GPUwerk has not run this model on two nodes; its one two-node run is a dense 32B model. Two cabled Sparks rent as a cluster for $1.79/hour; one Spark is $0.79/hour. The two-node cluster guide covers the setup.
Chat, single person
Pick a mixture-of-experts model with about 3B active parameters. Qwen3.6-35B-A3B at 86.3 tok/s and Qwen3-Coder-30B-A3B at 80.9 tok/s are both far above the roughly 10 tok/s line GPUwerk’s benchmarks page draws for a usable chat window. gpt-oss-120b is the larger option at 58.7 tok/s on llama.cpp build 7941 (llama.cpp DGX Spark results). Dense models are slow here: Nemotron Super 49B is at 5.79 tok/s and Llama 3.3 70B at 6.0.
Coding
Qwen3-Coder-30B-A3B is the coding pick GPUwerk has measured: 80.9 tok/s single-user and 5,705 tok/s prefill on an 8,192-token prompt, so a large file or diff as context is handled quickly. Qwen3.6-35B-A3B is faster in the reported runs. Qwen3.8-27B is the dense alternative if you want its quality and can accept 38 tok/s at short context.
Batch and agent-fleet work
Here aggregate throughput matters more than single-stream speed. gpt-oss-120b reaches 862.8 tok/s aggregate at 256 concurrent under vLLM. Nemotron Super 49B is slow for one user but reaches 695.1 tok/s at the same concurrency. GPUwerk measured Qwen3-Coder-30B-A3B at 3,235 tok/s total throughput at 256. Llama 3.3 70B peaks at 485 tok/s at 128 concurrent, then falls to 300 at 256, so keep batch jobs on it near 128.
Long prompts, short answers
Nemotron-3-Super-120B-A12B at NVFP4 decodes at 22.7 to 23.7 tok/s and prefills near 1,900 tok/s, which suits contract review and extraction work. It runs on stock vLLM without a fork. Decode at 4k, 50k and 100k tokens for the models above is compiled in DGX Spark long-context performance, and the newest large MoE releases across 2 to 8 Sparks in large MoE models on multiple Sparks.
When one Spark is not enough
Active parameters set speed on this hardware, and total parameters set fit. DeepSeek-V4-Flash has only 13B active, so it is fast enough, but its 284B total does not fit one Spark at the official precision. On one Spark it needs 2-bit or reduced builds, which sit below the 4-bit level where quality starts to drop on most models. A second Spark runs the official weights instead. Dense models past 30 to 50B hit the 273 GB/s bandwidth ceiling explained in the glossary, and a second Spark does not fix that for a single user.
How to pick
Start from the job. One person typing prompts: Qwen3.6-35B-A3B or Qwen3-Coder-30B-A3B. Many users or an agent fleet: gpt-oss-120b or Nemotron Super 49B, sized to your concurrency. Long documents with short answers: Nemotron-3-Super-120B. The official DeepSeek-V4-Flash weights: two Sparks.
Before you trust any number here, rent a DGX Spark at $0.79/hour, or two at $1.79/hour, billed per minute, and run your own prompts. A first engagement covers picking and tuning a model for a specific workload if you would rather hand that off.
FAQ
What is the best DGX Spark model for chat?
On one Spark, Qwen3.6-35B-A3B in NVFP4 with an MTP draft is the fastest current pick: SparkBench reports 86.3 tok/s single-stream at a 4k context fill (vLLM, 4 August 2026). Qwen3-Coder-30B-A3B is the fastest model GPUwerk has measured itself, at 80.9 tok/s on vLLM 0.27.1.
What is the best DGX Spark model for coding?
Qwen3.6-35B-A3B for speed on one Spark (86.3 tok/s reported at 4k). Qwen3-Coder-30B-A3B is the coding-tuned model GPUwerk measured at 80.9 tok/s, and it reached 3,235 tok/s total throughput at 256 concurrent requests. For a dense model, Qwen3.8-27B is reported at 38.1 tok/s at 4k with NVFP4 and speculative decoding on SGLang.
What does a second DGX Spark let me run?
Models whose official weights do not fit in 128 GB. DeepSeek-V4-Flash runs its official FP8 checkpoint on two Sparks with vLLM and TP=2 at about 44 tok/s single-stream (NVIDIA forum, May 2026). GPUwerk rents two cabled Sparks as a cluster for $1.79/hour.