DeepSeek V4-Flash-0731: The VRAM Reality
Official repo 166.9GB, GGUF tiers from IQ1 82.5GB to Q8 161.9GB, the 520GB claim checked against artifacts, and a pick-by-hardware table.
“V4-Flash needs 520GB” and “V4-Flash runs on a laptop” are both circulating this month, and both are wrong in instructive ways. The VRAM reality for the July 31 build is thinner than the hype: the official download is 166.9GB, the community quantization ladder runs from 82.5GB to 161.9GB, and nothing in the official artifacts comes near 520GB. This guide is part of the open-weight coding hub; official figures are marked official, community data is marked community, and the math behind each row is shown.
The official size figures
DeepSeek’s changelog (fetched August 12) describes the model as a 284B total / 13B active MoE with 1M context. Official Hugging Face metadata (fetched August 12) counts 304.2B total parameters — the lab’s “284B” and HF’s “304.2B” differ by counting convention (the lab’s figure excludes embeddings or auxiliary tensors; both are official, neither is wrong). The official config adds the serving-relevant details: 43 layers, 6 experts per token plus 1 shared, a single KV head (MLA-style), 1,048,576-token context.
The number that matters for hardware is the official repo listing, fetched today: 166.9GB across 48 weight files — about 0.55 bytes per parameter. DeepSeek ships this checkpoint in a compact precision, so the “native” download is not a float-weight monster; it is a Q6/Q8-class build straight from the vendor.
Quantization reality
| Build | Size | Source |
|---|---|---|
| Official HF repo | 166.9 GB | official (fetched 2026-08-12) |
| unsloth UD-Q8_K_XL | 161.9 GB | community (HF tree, fetched) |
| unsloth UD-Q4_K_XL | 155.1 GB | community |
| unsloth UD-IQ4_NL / IQ4_XS | 136.7 GB | community |
| unsloth UD-Q3_K_M / Q3_K_XL | 128.1 GB | community |
| unsloth UD-IQ3_S | 116.1 GB | community |
| unsloth UD-IQ2_M / IQ2_XXS | 90.9 GB | community |
| unsloth UD-IQ1_S | 82.5 GB | community |
The “520GB native” figure that circulates in threads: the official repo is 166.9GB, and a full BF16 conversion of 304.2B parameters would be ~610GB (2 bytes × 304.2B params — derived from official metadata). 520GB matches no official artifact and no standard precision of this model; treat it as unsourced until someone publishes the file listing behind it.
What the community actually runs (community data, not facts)
Field reports are labeled as what they are — community, not official — and the evidence base is the repos themselves: unsloth’s GGUF build has ~208K downloads and bartowski’s another ~55K (official HF download counters, fetched August 12), with an abliterated variant at ~153K. The realistic community consensus pattern is single-node, 192GB-class serving of the Q3_K_M/Q4 tiers, not single-GPU: the smallest build (IQ1_S, 82.5GB) already exceeds one 80GB GPU’s usable capacity once the runtime, KV cache and OS overhead are counted. Nothing official states a minimum GPU count — every “it runs on X GPUs” claim is community evidence.
Pick by hardware
Derived from the table above: VRAM needed ≈ weights + ~10–15% runtime headroom, and the 1M-token context pushes cache into the tens of GB for long prompts (MLA’s single KV head keeps it sub-quadratic — no official cache figure exists).
| Hardware | What fits | Verdict |
|---|---|---|
| Single 24–48GB GPU | nothing (min 82.5GB) | out |
| Single 80GB GPU | nothing (82.5GB + overhead) | out |
| 2×80GB (160GB) | Q3_K_M 128.1GB and below | the community sweet spot |
| 2×96GB / 4×48GB (192GB) | Q4_K_XL, Q8, official 166.9GB | recommended |
| 4×80GB (320GB) | everything on the ladder | comfortable |
| 8×80GB (640GB) | theoretical BF16 (~610GB) | only if you truly need it |
Compute is not the constraint — 13B active parameters make per-token cost small (official). Memory is the entire game, and 192GB-class nodes are the pragmatic floor for the official build. For the budget case, the price-quality comparison shows why self-hosting may not beat the API at all below that line.
Official source
- Official repo + file listing: deepseek-ai/DeepSeek-V4-Flash-0731 (metadata and tree fetched 2026-08-12)
- Release card: api-docs.deepseek.com/updates (fetched 2026-08-12)
- Community builds: unsloth/DeepSeek-V4-Flash-0731-GGUF (tree fetched 2026-08-12)
- Sibling coverage: V4-Flash 0731 vs Max, cost deep-dive, pricing reconciliation
First published August 12, 2026. Official figures from the pages listed and fetched this session; community rows are labeled community; derived rows show their inputs.