Docs/Models/DeepSeek-V4-Flash
Models

DeepSeek-V4-Flash on DGX Spark: 1 vs 2 nodes, tok/s

Samuel Seidel · Published September 7, 2026 · Updated 24 October 2026

Two DGX Sparks run the official FP8 DeepSeek-V4-Flash at about 44 tok/s single-stream (vLLM, TP=2, 200K context, NVIDIA forum, 16 May 2026). One Spark needs a 2-bit or reduced build and is reported at 13 to 55 tok/s, depending on engine and quantization. The model has 284B total parameters and 13B active per token (model card, read 24 October 2026). Every figure below is reported by a third party. GPUwerk has not run this model on its fleet.

One Spark vs two Sparks

Setup Engine / image Quant Context Decode tok/s, single stream Source Date
1 SparkNVFP4 build; engine not namedNVFP4, reducedtrimmed~55 typical, 78 peakopenclawdc (compiled from community recipes)2026-08-05
1 Sparkds4 fork for GB10 (Entrpi, based on antirez’s ds4), speculative decoding2-bit, antirez’s 0731 quant12k in use~28 (~22 at 240k; 59 aggregate at 12 concurrent)NVIDIA forum2026-08-01
1 SparkDwarfStar 4 (antirez’s ds4)asymmetric GGUF, ~110 GiB1M configured26.7noze, citing MiaAI-Lab2026-08-06
1 SparkvLLM-Moet-GB10 (patched vLLM fork), ASUS Ascent GX102-bit expert planes + FP4 tiernot stated~13 prose / ~21 code, prefill ~753Hugging Face thread2026-08-06 to 08-09
2 Sparks, TP=2vLLM (jasl fork, commit dda4668b) built with eugr’s spark-vllm-docker, MTP 2official FP8, FP8 KV cache200K~44 (advantage) (~45 aggregate at 2 concurrent)NVIDIA forum2026-05-16
2 Sparks, TP=2vLLM 0.21.1, aidendle94/sparkrun-vllm-ds4-gb10, MTP 2official checkpoint, FP8 KV cache1M45.7 at empty context, 30.4 at 980KNVIDIA forum2026-06-04
2 Sparks, TP=2vLLM, tensor parallel over RoCE0731 checkpoint, quant not stated1M82.4 (2,048-token completions, short prompts)noze, citing MiaAI-Lab2026-08-06
2 Sparksnot statednot stated256K asked“around 25”NVIDIA forum2026-05-21

Sources read 24 October 2026. “0731” is the July 2026 DeepSeek-V4-Flash-0731 checkpoint; the other rows use the original release. Speculative decoding (MTP) is on in most rows, so the figures depend on how predictable the text is: the Hugging Face thread reports ~13 tok/s on prose at temperature 1.0 and ~21 on code at temperature 0 with the same build. An earlier post in that thread gave 14.3 tok/s; the newest post says ~13, and this page and the benchmarks page both use ~13. Four Sparks with TP=4 are reported at about 70 tok/s single-stream on official FP8 (same forum thread, post 174, 3 June 2026).

What the second Spark buys

It buys the official weights. The vLLM recipe (updated 29 September 2026, read 24 October 2026) lists the model as 284B total and 13B active, with FP4 MoE experts and FP8 for attention, norm and router weights. Its DGX Spark (GB10) section uses two nodes with --tensor-parallel-size 2 and the community eugr/spark-vllm-b12x image, and says one GB10’s 128 GB is below the FP8 and NVFP4 checkpoint footprint. One Spark only runs the model after dropping to 2-bit experts or a reduced NVFP4 build with trimmed context.

The speed gap is smaller than the quality gap. The best one-Spark report (about 55 tok/s, reduced NVFP4) is faster than the typical two-Spark FP8 run (about 44 tok/s). What two Sparks add is the official checkpoint and long context: the June run held 30 tok/s at 980K tokens.

Memory footprint

The model card lists the checkpoint as mixed FP4 and FP8 and does not give a size in GB. The vLLM recipe says that checkpoint does not fit in one GB10’s 128 GB. The ds4 route on one Spark uses an asymmetric GGUF of about 110 GiB of weights, per the noze summary of MiaAI-Lab’s setup. On two Sparks the weights are split across both nodes. GPUwerk’s cluster gives 256 GB in total across two nodes, not a single memory pool, so the runtime must support distributed execution.

How to run it

On two Sparks, follow the DGX Spark section of the vLLM recipe. It needs the community Spark vLLM build rather than the stock vllm/vllm-openai image, plus the ConnectX interface names and head-node IP on each node. The forum threads above give full commands for their own setups. GPUwerk has not verified a command for this model, so this page links them rather than copying one.

On one Spark, the ds4 path is the documented one: antirez’s ds4 lists DGX Spark as its main CUDA target, and the GB10 fork’s installer is at Entrpi/ds4-on-spark. Both use 2-bit expert quants.

When this model is the wrong choice

On one Spark, the 2-bit and reduced builds trade quality for fit. They sit below the 4-bit level where GPUwerk’s quantization guidance says quality starts to drop on most models. Every working setup above also depends on a fork or a community image, not a stock release. If you need a model that fits one Spark at 4-bit or better on stock vLLM, pick from the models page, for example Nemotron-3-Super-120B at NVFP4.

Try it

One Spark is $0.79/hour. Two Sparks, cabled as a cluster, are $1.79/hour. Both bill per minute against a prepaid balance (see the pricing page). The two-Spark cluster is the setup the vLLM recipe targets. Bring the recipe and check the numbers on your own prompts.

Comparing with your own GPU or Mac? FitLocally's page for DeepSeek-V4-Flash lists which graphics cards, Macs and mini PCs fit it at each quantization, with estimated speeds. FitLocally is a separate site I publish; its DGX Spark figures come from these guides.

FAQ

Does DeepSeek-V4-Flash fit on one DGX Spark?

Not at its official precision. The vLLM recipe says one GB10's 128 GB is below the FP8 and NVFP4 checkpoint footprint and uses two Sparks with TP=2. One Spark works only with a smaller build: 2-bit expert quants on ds4 or a patched vLLM fork, or a reduced NVFP4 build with trimmed context.

How fast is DeepSeek-V4-Flash on one Spark?

Reported single-stream decode on one Spark runs from about 13 tok/s (patched vLLM fork, 2-bit experts, prose) to about 55 tok/s typical (NVFP4, reduced context). The ds4 route is reported at about 26 to 28 tok/s. GPUwerk has not reproduced these numbers.

How fast is DeepSeek-V4-Flash on two Sparks?

About 44 tok/s single-stream on the official FP8 checkpoint with vLLM, TP=2 and MTP at 200K context (NVIDIA forum, May 2026). A June 2026 run reports 45.7 tok/s at empty context and 30.4 tok/s at 980K. Two linked Sparks rent from GPUwerk as a cluster for $1.79/hour.

See alsoFull benchmark numbers across engines See alsoWhich models fit in 128 GB
First top-up: pay $20, get $30 in credit

Run it yourself, not our numbers.

A dedicated Spark, deployed in minutes.

Create account