I think people with 4 DGX Spark are good to go. Rest of us needs to wait for quantized version. DeepSeek-V4 - a deepseek-ai Collection We’re on a journey to advance and democratize artificial intelligence through open source and open science.

Sorry, I didn’t explain it clearly earlier. I’ve been testing DeepSeek‑V4 on a dual‑DGX Spark setup, and when I follow your parameter settings I can reproduce the same performance as you. However, when ISL > 128 or concurrency > 4, the benchmark fails and the Dual Ray compiled DAG runs into issues. …

Can you give me a reproduce steps like CURL command? I know the tool calling has issues, but I don’t have samples to make a proper fix. PS: I’m rewriting essential kernels in the vLLM side, to avoid DeepGEMM, so it maybe unstable in a short-term. vLLM team has reached to me that says they want to …

Thanks by the way for your work. It’s very much appreciated!

I’m rewriting kernels to avoid DeepGEMM, so you may encounter crashes and instability. Please share your error log and VLLM command with me, so I can help test. I only have 1 spark, so I can’t fully test on GB10, my main dev machine is 2x RTX Pro 6000

I added VLLM_DEEPSEEK_V4_USE_DEEPGEMM_SM12X_KERNELS=1 to the ds4-sm120-full branch. I still wish you could try the default new kernel and let me know if you run into any errors or issues. You can switch to the previous behavior if the new kernel is blocking your job.

could you try with this? when i removed active memory, which injects these xml tags, it got little more stable https://gist.github.com/teason2021/869b67f636bac3bcbfc8c3b784075d54 deepseek_v4_flash.sh curl -X POST "https://your-endpoint.example.com/v1/chat/completions" \ -H "Content-T…

I need to find an SM90/100 GPU server to do the correctness check. Give me a few days

[image] Teason2026: -apply-vllm-pr yes, if you check SparkRun https://sparkrun.dev there are models using sglang already. You can also see we have SGLang recipes for some models at https://spark-arena.com

DeepSeek-V4-Flash on dual DGX Spark (TP=2) — ToolCall-15 Score: 90/100 (27/30) ★★★★★ Excellent Hardware: 2 × DGX Spark (GB10, 128 GB unified, sm_121, 200 Gb ConnectX-7 RoCE), torchrun --no-ray TP=2, ascent head (10.1.0.2) + spark worker. Image: deepseek (built from spark-vllm-docker --apply-vllm-p…

Deepseek V4 released

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

arthurdroz April 29, 2026, 12:43pm 89

I created an experimental eugr recipe PR here, seeing some issues with tool calling but getting around 22 t/s Add DeepSeek V4 Flash recipe by arthur-drozdov · Pull Request #219 · eugr/spark-vllm-docker · GitHub

Topic		Replies	Views
DeepSeek-V4-Flash (official FP8) running across 2x DGX Spark — TP=2, MTP, 200K ctx, recipe + numbers DGX Spark / GB10 deepseek	261	32227	August 21, 2026
DeepSeekV4-Flash hybrid quant, 1x DGX Spark: antirez's optimized 128 GB MLX recipe ported to vLLM for GB10 DGX Spark / GB10 Projects deepseek	18	2849	May 11, 2026
Deepseek v4 Flash on 2 Nodes DGX Spark / GB10 Projects deepseek	71	9297	June 15, 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	52116	September 10, 2026
Official NVidia optimized DeepSeek-V4-Flash models? DGX Spark / GB10 deepseek	28	3814	July 11, 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	7321	September 5, 2026
DeepSeekv4 Flash for 1x Spark [REAP25] [PrismaAURA] DGX Spark / GB10 Projects llama , agentic-ai , deepseek	3	1123	July 19, 2026
DeepSeek-V4-Flash-0731-NVFP4 DGX Spark / GB10 deepseek	9	1098	September 3, 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	1824	June 19, 2026
How can I get DeepSeek-V4-Pro running on my 8 Spark cluster? DGX Spark / GB10 deepseek	27	1360	July 7, 2026

Deepseek V4 released

Related topics