RUN THIS LLM
Search local LLM hardware requirements
DeepSeek V4 Flash 0731
DeepSeek · 284B MoE · General
Official release of DeepSeek V4 Flash superseding the preview: same 284B MoE / 13B active architecture with hybrid compressed sparse attention and a built-in DSpark speculative-decoding head, but far stronger agentic coding (82.7 Terminal-Bench 2.1, 54.4 DeepSWE). Text-only, 1M context, MIT license.
VRAM Requirements
| Quantization | VRAM |
|---|---|
| Q4_K_M (smallest) | 170.4 GB |
| Q8_0 (balanced) | 312.4 GB |
| FP16 (full quality) | 568 GB |
Specifications
- Parameters: 284B MoE (13B active per token)
- Category: General
- Max context: 1M tokens
- System RAM: 192 GB minimum
- HuggingFace: deepseek-ai/DeepSeek-V4-Flash-0731
Benchmarks
- GPQA: 90 — Graduate-level science questions (Diamond)
Loading interactive analysis...