deepseek-ai · deepseek_v4
DeepSeek-V4-Flash
DeepSeek-V4-Flash by deepseek-ai: a 290.9B-parameter mixture-of-experts open-weight model under the mit license.
1,590,858 downloads · 2,236 likes · Model card · synced 2026-09-19
Specifications
- Parameters
- 290.9B (6.8B active)
- Architecture
- moe · deepseek_v4
- Layers
- 43
- Hidden size
- 4,096
- KV heads · head dim
- 1 · 512
- Vocabulary
- 129,280
- Native dtype
- bfloat16
- Context window
- 1,048,576 tokens
- Released
- 2026-04-22
- Last modified on Hub
- 2026-06-22
Features & licensing
- License
- mit
- Commercial use
- Yes
- Modalities
- text
- Tool / function calling
- No
- Reasoning mode
- No
- Languages
- EN
- Quantised variants on Hub
- GGUF, bitsandbytes
- Pipeline
- text-generation
Task fit
Editorial scores (0–100) used by the recommendation engine.
- Chat70
- RAG / Q&A70
- Code70
- Summarisation70
- Extraction70
- Agents70
Hardware to run DeepSeek-V4-Flash
Weights need about 640 GB at FP16, 336 GB at INT8 and 192 GB at INT4. The KV cache adds roughly 88 MB per 1,000 tokens per request. For a reference workload of 5 requests per second with 1,500 input and 300 output tokens, total GPU memory is around 257.8 GB, which fits on 4 × NVIDIA A100 (80 GB).
| Configuration | VRAM | Utilisation | Est. first token | Cloud / month |
|---|---|---|---|---|
| 6 × NVIDIA L40S (48 GB) | 288 GB | 90 % | ~0.2 s | $8,322 |
| 4 × NVIDIA A100 (80 GB)recommended | 320 GB | 81 % | ~0.1 s | $9,344 |
| 4 × NVIDIA H100 (80 GB) | 320 GB | 81 % | ~0.0 s | $13,140 |
Estimates only. Run an assessment for your own traffic.
Auto-imported from Hugging Face. Architecture, license and feature data are synced from the model repository; the task-fit scores below are provisional defaults until a curator reviews them.