DeepSeek
DeepSeek-V4-Flash
A 284B mixture-of-experts model. Official FP8 needs two linked Sparks.
Recommended: B300, 288 GB HBM3e, $6.49 / hour, billed per minute. Check availability.
Configure
Every build below is published on Hugging Face and fits at least one GPU we rent. Pick the quantization and engine: the memory it needs, the GPUs it fits and the recommended GPU follow, and Deploy takes that exact build, revision and GPU to the console. Only the one-click preset is tested by GPUwerk.
| Quantization | FP8 |
|---|---|
| Engine | vLLM |
| Status | Untested by GPUwerk |
| Weights | 159.6 GB |
| GPU memory it needs | About 186.5 GB with a 32k-token cache, an estimate |
| Recommended GPU | B300, $6.49 / hour |
| Weights source | deepseek-ai/DeepSeek-V4-Flash, revision 60d8d70770c6 |
| Licence on the card | mit |
| GPU | Memory | Price | Availability | Deploy |
|---|---|---|---|---|
| RTX 5090x86 · Blackwell | 32 GB GDDR7 | Does not fit: needs about 186 GB of GPU memory (weights 159.6 GB plus cache, an estimate); RTX 5090 offers 29 GB to it. | ||
| DGX SparkArm · GB10 | 128 GB unified | Does not fit: needs about 186 GB of GPU memory (weights 159.6 GB plus cache, an estimate); DGX Spark offers 96 GB to it. | ||
| RTX PRO 6000x86 · Blackwell | 96 GB GDDR7 | Does not fit: needs about 186 GB of GPU memory (weights 159.6 GB plus cache, an estimate); RTX PRO 6000 offers 86 GB to it. | ||
| H100 SXMx86 · Hopper | 80 GB HBM3 | Does not fit: needs about 186 GB of GPU memory (weights 159.6 GB plus cache, an estimate); H100 offers 72 GB to it. | ||
| H200x86 · Hopper | 141 GB HBM3e | Does not fit: needs about 186 GB of GPU memory (weights 159.6 GB plus cache, an estimate); H200 offers 127 GB to it. | ||
| B200x86 · Blackwell | 180 GB HBM3e | Does not fit: needs about 186 GB of GPU memory (weights 159.6 GB plus cache, an estimate); B200 offers 162 GB to it. | ||
| B300x86 · Blackwell · recommended | 288 GB HBM3e | $6.49 / hour | Check availability | Run it yourself DeepSeek-V4-Flash FP8 on B300 |
Builds by the model's maker, NVIDIA and Unsloth, read from Hugging Face on 21 November 2026. Memory is an estimate (weights x 1.15 plus 3 GB); GGUF runs on llama.cpp, which GPUwerk builds for the DGX Spark only, and NVFP4 needs a Blackwell GPU. SGLang is not offered: there is no SGLang runtime yet.
Details
About the model.
| Maker | DeepSeek |
|---|---|
| Task | Text |
| Parameters | 284B total, 13B active (mixture of experts) |
| Licence | See the model card |
No speed figure here: GPUwerk lists only figures it measured itself. Reported figures, with their sources, are in the model guide.
Quick start
- Deploy a 2 × DGX Spark with a base image (DGX OS or Ubuntu).
- Follow the DeepSeek-V4-Flash guide to download the weights and start vLLM.
- Connect over SSH or the browser terminal; the machine is yours until you terminate it.
SSH, the browser terminal, VNC and your workspace work the same on every instance. Terminate when you are done and billing ends.
Related models
Qwen3 Coder 30B
A coding model for agents and editors, with tool calls turned on.
Qwen3.6 35B-A3B
A fast mixture-of-experts model for chat and agents, also on the pay-per-token API.
GPT-OSS 120B
OpenAI's open-weight 120B model, quantized to fit one DGX Spark.