RUN THIS LLM
Search local LLM hardware requirements
DeepSeek V4.1 Flash 552B MoE
DeepSeek · 552B MoE · Vision
Multimodal 552B MoE with 8-16B active, built as a 20-layer causal encoder + 20-layer decoder (CED) with Compressed Sparse Attention 2 that shares KV across layers — global KV cache per token is ~4x smaller than V4 Flash. 1M context, MIT license; also carries a 196B sparsely-accessed Engram memory table.
VRAM Requirements
| Quantization | VRAM |
|---|---|
| Q4_K_M (smallest) | 331.2 GB |
| Q8_0 (balanced) | 607.2 GB |
| FP16 (full quality) | 1104 GB |
Specifications
- Parameters: 552B MoE (16B active per token)
- Category: Vision
- Max context: 1M tokens
- System RAM: 384 GB minimum
- HuggingFace: deepseek-ai/DeepSeek-V4.1-Flash
Benchmarks
- GPQA: 91 — Graduate-level science questions (Diamond)
Loading interactive analysis...