DeepSeek · Released April 2026
DeepSeek V4 Flash
What it takes to run DeepSeek V4 Flash locally: memory by quantization, the GPUs that handle it, and how fast it runs.
Updated
- Parameters
- 284B (13B active)
- Architecture
- Mixture of experts
- Context
- 1.0M tokens
- License
- MIT license
- Memory at FP4
- ~168GB
What do you need to run DeepSeek V4 Flash?
DeepSeek V4 Flash needs about 168GB of VRAM at FP4 with 8K context (162GB for the weights plus KV cache and overhead). Because it is a mixture-of-experts model with 13B active parameters, it generates text much faster than a dense model of the same size, and it tolerates partial CPU offload better.
What to buy
GPUs for DeepSeek V4 Flash
Speeds are estimates at the quantization shown with 8K context. We may earn a commission when you buy through these links. Learn more.
Memory
DeepSeek V4 Flash VRAM requirements by quantization
DeepSeek V4 Flash ships natively in FP4, so there is no need to pick a quantization. Totals include the KV cache for the context length shown plus runtime overhead.
| Quantization | Quality | Weights | Total @ 8K ctx | Total @ 32K ctx |
|---|---|---|---|---|
| FP4 | NativeDefault | 162 GB | 168 GB | 168 GB |
KV cache note: Compressed attention: layers alternate between compressed sparse attention (about 4 tokens merged per cache entry) and heavily compressed attention (about 128 tokens per entry), so the cache grows far more slowly than standard or MLA attention. DeepSeek's published figures put a full 1M-token cache at roughly 5GB, about 5,000 bytes per token. Treat this as approximate; actual use depends on runtime support.
Performance
How fast DeepSeek V4 Flash runs
Single-user generation speed on popular GPUs that fit the model entirely in memory, at the best quantization each can hold.
Overview
About DeepSeek V4 Flash
DeepSeek V4 Flash is the smaller model of DeepSeek's April 2026 V4 generation: 284B total parameters, 13B active, a 1M-token context, and a toggleable thinking mode, refreshed with improved 0731 weights in July. Its compressed attention makes long contexts remarkably cheap, and its expert weights ship natively in FP4, so the full-quality build is only about 162GB. Running it at full quality takes a 256GB Mac Studio or a multi-GPU server with about 170GB of memory; 3-bit dynamic builds of roughly 103GB to 110GB fit 128GB machines with some quality loss. DeepSeek's September V4.1 Flash is larger (552B, including 196B of Engram memory), so V4 Flash remains the easier DeepSeek to run locally.
Strengths
- +Near-frontier reasoning and coding with only 13B active parameters
- +1M-token context with a very small KV cache
- +Native FP4 experts: the full-quality build is about 162GB
- +MIT license
Weaknesses
- −Needs about 170GB for full quality; 128GB machines must drop to 3-bit
- −Text only (vision is in a separate experimental release)
- −Runtime support for its compressed attention arrived months after launch and is still maturing
Setup
How to run DeepSeek V4 Flash
$ llama-server -hf unsloth/DeepSeek-V4-Flash-GGUF -c 8192 -ngl 99Weights: deepseek-ai/DeepSeek-V4-Flash · GGUF: unsloth/DeepSeek-V4-Flash-GGUF
Alternatives
Models like DeepSeek V4 Flash
Z.ai
GLM-5.3-Flash
320B (18B active) · 1.0M ctx
Alibaba Qwen
Qwen3.5 122B-A10B
122B (10B active) · 256K ctx
DeepSeek
DeepSeek R1 (0528)
671B (37B active) · 128K ctx
Z.ai
GLM-5.3
744B (40B active) · 1.0M ctx
Frequently asked questions
How much VRAM do you need to run DeepSeek V4 Flash?
About 168GB at FP4 with 8K tokens of context: 162.0GB of weights, 0.0GB of KV cache, and 5.5GB of runtime overhead. Higher-precision quants and longer contexts need more; see the table above.
Is DeepSeek V4 Flash free for commercial use?
DeepSeek V4 Flash is released under the MIT license. That is a permissive license, so commercial use is allowed.
How long a context does DeepSeek V4 Flash support?
Up to 1,048,576 tokens (1.0M). Long contexts need a lot of extra memory for the KV cache, so most local setups run 8K to 32K.
Sources: DeepSeek V4 Flash model card (Hugging Face), Hugging Face: DeepSeek-V4, a million-token context that agents can actually use, Unsloth: How to run DeepSeek-V4 locally. Capability score basis: Artificial Analysis Intelligence Index: 50 for the 0731 refresh before the v4.3 rescale (August 2026), close to Qwen3.8 27B; placed slightly above it for broader knowledge and below GLM-5.3-Flash (42 on v4.3, September 2026).