About DeepSeek V4 Flash
DeepSeek V4 Flash is the efficiency-focused member of DeepSeek's V4 series: an MIT-licensed open-weight mixture-of-experts model whose active parameters per token are a small slice of its large total, paired with heavily compressed attention that keeps its million-token context practical. DeepSeek pitches it at fast, intuitive responses for routine work, with switchable thinking modes when a task needs actual deliberation.
It leads a double life by design — the same weights serve DeepSeek's low-cost API and anyone's own infrastructure — which is why it appears in both our cost and VRAM guides.
DeepSeek V4 Flash is DeepSeek's 284B-parameter model released in April 2026, with a 1024K-token context window. It is a mixture-of-experts model: all 284B parameters must sit in memory, but only ~13B are active per token, which is what makes it fast for its size. It uses a latent-attention (MLA) design that compresses its KV cache, keeping even very long contexts cheap in memory: its KV cache is about 0.4 GB at an 8K context, 6.5 GB at 128K, and 51.9 GB at the full 1024K window (FP16 cache).
Running it yourself is a serious-hardware proposition despite the friendly API pricing: mixture-of-experts means the entire expert pool must sit in memory, so the totals above track the full parameter count, well beyond single consumer cards even heavily quantised. Where it fits — multi-GPU servers, the largest unified-memory machines — the compressed attention design pays off unusually well: the KV cache barely grows with context, so the million-token window costs almost nothing beyond the weights. For most individuals the honest answer is that the API is the way to use this model, and the local option is what the licence makes possible.
For most people Q4_K_M is the sweet spot — the most popular quality/size trade-off — while Q8 is near-lossless if you have the memory. Totals above include the KV cache and a realistic framework overhead, so they are what you should expect to see in practice rather than just the download size. Weight sizes are calibrated against real GGUF files — see the methodology.
Frequently asked questions
How much VRAM does DeepSeek V4 Flash need?
At Q4_K_M with an 8K context, DeepSeek V4 Flash needs about 192 GB (weights 173 GB + KV cache + overhead). The smallest common hardware that fits is a Apple Silicon 512GB unified.
Can an RTX 4090 (24GB) run DeepSeek V4 Flash?
Not fully in VRAM. DeepSeek V4 Flash needs about 192 GB even at Q4_K_M, so a 24 GB card would have to offload layers to system RAM at a large speed penalty.
Can a Mac run DeepSeek V4 Flash?
Yes — Apple Silicon with 512 GB of unified memory or more (macOS lets the GPU use ~75% of it, ~384 GB) runs DeepSeek V4 Flash at Q4_K_M.
How much VRAM does DeepSeek V4 Flash need at its full 1024K context?
The KV cache grows with context: about 0.4 GB at 8K tokens versus 51.9 GB at the full 1024K window (FP16 cache). Add that to the weights (173 GB at Q4_K_M) plus overhead — long contexts can cost more than a whole quantisation step.
Related
VRAM calculator for any model ·Token counter · DeepSeek V4 Flash API pricing
Similar-size local models: Qwen3.5 397B-A17B (397B) · Llama 4 Maverick (400B) · Mistral Large 3 (675B) · Mistral Small 4 (119B) · GPT-OSS 120B (117B)
Last updated 2026-09-03. Architecture figures from the model's published config.json; see the methodology.