Got the NEW Deepseek up and running. There was a small config change you have t do to get Acceptence back up. For how it ships it will hit 30 tok/s but with the config change it brought acceptence back to its normal speeds as preview. AGENT- 1/ 🔍 Upgraded to the official DeepSeek-V4-Flash-0731 on …

WELL DONE Tony! finally DS4F is usable :)

Amazing, thanks! Will try it the upcoming days!

If we do not use NVFP4 KV cache, how much KV cache size can we get with this? Approximately half? This is impressive, and thanks for your work! Nvidia really has to thank Deepseek, as I just ordered a new Spark because of this new release 😂

Hey Tony, is this expected: => => naming to Docker Hub Container Image Library | App Containerization 0.0s INFO 08-01 14:48:04 [importing.py:46] Triton is installed but 0 active driver(s) found (expected 1). Disabling Triton to prevent runtime errors. INFO 08-01 …

I have no clue what that is brother. Did you ask your agent

It’s from the docker build. I’ll ask my agent now that the build succeeded (and the agent has a brain again), but I wondered if you saw similar output on your docker and just ignored

I got this going, thanks! Haven’t tested the upper realms beyond 300k yet, but it seems to work like the older dspark setup I had going before, maybe a touch smarter and faster even, with larger context room. All good things.

@tonyd615 Very curious how you test the token generation speed. Also what concurrency and depth are you testing at. With eugr/llama-benchy i get different results (lower) than those you published. Thanks.

Thanks Tony for sharing the recipe. It actually worked for me on my system too. However, NVFP4 for KV-cache didn’t work for me, despite incorporating your patches. Also only getting 12tok/s because >90% of draft tokens are being rejected. Here are the exact measured performance results from the l…

DeepSeek-v4-Flash-0731-DSpark-1M-NVFP4-KV-2x-DGX-Spark

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

ammarabbaxi13 August 18, 2026, 7:00am 15

Hi tony, on vllm 0.21. I’m getting errors with claude code, I heard that was fixed with vllm 0.23. Is there any way I can upgrade to vllm 0.23 to run this

Topic		Replies	Views
Official NVidia optimized DeepSeek-V4-Flash models? DGX Spark / GB10 deepseek	28	3474	July 11, 2026
DeepSeek-V4-Flash-DSpark on 2× DGX Spark (GB10) — big single-stream speed boost (~60-67 tok/s) + 1M context, now with concurrency DGX Spark / GB10 deepseek	102	11473	August 7, 2026
Instructions for running Deepseek-v4-flash with DSpark using Eugr's repo DGX Spark / GB10 Projects deepseek	99	9149	August 23, 2026
DeepSeek-V4-Flash (official FP8) running across 2x DGX Spark — TP=2, MTP, 200K ctx, recipe + numbers DGX Spark / GB10 deepseek	261	30684	August 21, 2026
Deepseek v4 Flash on 2 Nodes DGX Spark / GB10 Projects deepseek	71	8777	June 15, 2026
1x Spark tuned DSpark for DeepSeek-V4-Flash: 35 tok/s, 800+ prefill, and fast multi-agent serving DGX Spark / GB10 Projects deepseek	161	15415	August 19, 2026
New DeepSeek-V4-Flash-DSpark DGX Spark / GB10 deepseek	5	5056	June 29, 2026
DeepSeek-V4-Flash on 4× DGX Spark via vLLM (jasl fork, TP=4, RDMA, MTP) — 49–54 tok/s single-stream, full recipe + the traps DGX Spark / GB10 Projects deepseek	3	1594	June 19, 2026
DeepSeek v4 Flash (Aiden Recipe from Reddit) - 1M token session operational, Cuda 12.1 tailored for DGX Spark GB10 DGX Spark / GB10 deepseek	823	48559	September 10, 2026
C1 1058pp/s, 52 tg/s on 1x dgx spark on Deepseek V4 flash 0731 full 256 experts DGX Spark / GB10 Projects deepseek	94	6670	September 5, 2026

DeepSeek-v4-Flash-0731-DSpark-1M-NVFP4-KV-2x-DGX-Spark

Related topics