DeepSeek v4 Flash Vision Exp is Released as Open Weights

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10
450

Finally got my deployment put together in a form that can be shared.

GitHub - bumblebeers/DSv4-Flash-Vision-Exp-on-2x-Sparks: Serve DeepSeek-V4-Flash-Vision-Exp (305B, 1M context, native vision) on 2x NVIDIA DGX Spark at TP=2. Reproducible deployment: pinned image, serving config, gated launcher, verification battery, benchmarks. · GitHub.

This is an agent-authored repo (i.e., it’s sloppy), but you should be able to point your own agent at it and have it recreate the stack.

The Headlines:

  • 4.5M+ KV pool (fp8)
  • ~40 tk/s decode on average (breaks 60 tk/s on structured/code gen)
  • Good concurrency 25-30 tk/s @ 6 concurrent pelican requests with no quality degradation.
  • MTP = 3 (high total acceptance)
  • Image coordinate fix included
  • built on b12x (the 0rand image)

Speed

Depth Prefill tok/s Decode tok/s
0 2,596 41.1
4,096 2,035 45.4
16,384 1,931 40.3
32,768 1,871 43.1
65,536 1,803 42.1
131,072 1,682 38.8
262,144 1,527 43.1
524,288 1,244 34.9

Pelican Bench

Tool-Eval-Bench @ C=12

Tool-Call Benchmark — deepseek-ai/DeepSeek-V4-Flash-Vision-Exp

  • Run ID: 2026-09-28T18-04-16.357322Z_cf83109d
  • Date: 2026-09-28T18:14:25.145926+00:00
  • tool-eval-bench: v2.7.1.dev7+gbd35ba91b
  • Final Score: 90 / 100
  • Total Points: 165 / 184
  • Rating: ★★★★★ Excellent
Run Context
Parameter Value
Backend vllm
Server http://***:8000/v1
Model (API) deepseek-ai/DeepSeek-V4-Flash-Vision-Exp
Temperature 0.0
Seed —
Max Turns 8
Timeout 120.0s
Scenarios all (92)
Parallel 12
Error Rate 0.0
Thinking enabled
Max Tokens 16384

Feedback

I’ll likely be running this stack for a while — until there is a better model that can run at this speed at native precision. So if you notice any issues, or see room for improvement let me know and I’ll get it in the repo.

451

Is the @OllieO stack the top pick now?

452

Honestly, I think yes. I tried use mainline with couple of fixes from OllieO’s stack suggested above but found it still not as polished even if it seemed a bit more performant.

453

I’ve been a little out of the loop (just using DSV4F happily on my sparks getting work done) but I’m curious if tensorfold will bring a performance improvement to dsv4f like it seems to with GLM 5.3 Flash. I really like Deepseek for my uses, which are less code heavy and more infrastructure / assistant work.

454

Do you have 2 or 4? If 4 go with DSv4.1F for twice the performance and 550B parameters of smarts.

456

Unfortunately only 2 😭

457

Yeah - before I switched to Mia’s tp4 I was running 2x2 of Ollie’s stack for several weeks. Best I tested, for sure.

458

I still really like this model. Anyone who is running dual sparks should still consider DeepSeek V4 Flash Exp Vision and shoot it out against other options for their workflow. Switching back to it after a week of “cutting edge” GLM 5.3 Flash recipes and it is cranking through work GLM messed up in my home lab.

459

With proper guardrails qwen3.8 fn ,dsv4 vision exp, glm flash all work fine for me either coding or building servers task.

Weakness of dsv4 vision is its vision capability. Other than that it’s just fine.

460

Indeed! I need to try Qwen but we are truly “eating good.” I just wanted to show love for a model that I don’t read about as often but I personally fine very relevant for a dual spark cluster for my workflow. If this was that I could ever get from the sparks I would be satisfied honestly, but of course I’ll be trying to get more performance out of these little boxes soon after I’ve built more infra.