Overview
DeepSeek V4 Flash VRAM requirements start at about 160 GB for the weights alone, because the official Hugging Face repo is 160 GB, so plan on roughly 161 to 162 GB with the site’s assumed 1 to 2 GB overhead, before any KV cache. That is a 284B-parameter mixture-of-experts model with 13B parameters active per token, stored as FP4 experts and FP8 everything else. The older DeepSeek R1 671B is 689 GB in its repo; the R1 distills (70B, 32B, 14B, 8B) are ordinary dense models you can size with the formula on methodology. The V4 Flash KV cache is unknown here, and tokens per second are unknown everywhere on this page. Prices: check current retailer price.
What do the model cards say?
| Model | Total / active params | Weight format | Context | License | Repo size |
|---|---|---|---|---|---|
| DeepSeek-V4-Flash | 284B / 13B (safetensors metadata: 291B) | FP4 expert weights, FP8 for most other parameters | 1M tokens | MIT | 160 GB, 46 shards |
| DeepSeek-R1 | 671B / 37B (safetensors metadata: 684B) | FP8 (tensor types BF16, F8_E4M3, F32) | 128K | MIT | 689 GB, 163 shards |
Neither card states a memory requirement. The V4 Flash card’s only memory guidance is that Think Max mode wants a context window of at least 384K tokens. Its config.json lists 43 layers, 256 routed experts with 6 per token, and a per-layer KV compression schedule.
What are the DeepSeek V4 Flash VRAM requirements?
The formula is weights_GB = params × bits ÷ 8 ÷ 1e9 (llm-quantization-explained). Applied to the card’s 284B:
| If every weight were | Weights (284B) | Weights (291B metadata count) |
|---|---|---|
| 2 bits | 71.0 GB | 72.8 GB |
| 4 bits (FP4) | 142.0 GB | 145.5 GB |
| 8 bits (FP8) | 284.0 GB | 291.0 GB |
The real repo is 160 GB, which sits between the 4-bit and 8-bit rows. That fits the card: experts are FP4, the rest FP8, and 160 GB × 8 ÷ 291B is 4.40 bits per parameter on average. LM Studio’s page lists a 156.40 GB GGUF download and says “you need at least 156 GB of RAM” for the smallest build; it does not name the quantization. The Ollama page lists no sizes and says the model “was retired on August 27, 2026”, so do not plan around an Ollama tag.
Add overhead (assumed 1 to 2 GB, not measured) and you get 161 to 162 GB. KV cache is not added because it is unknown: V4 Flash compresses its cache per layer (ratios of 4 and 128 in the config, plus a 128-token sliding window), so the plain KV formula on methodology does not describe it and the card gives no figure.
Why do sources say 82.5 GB to 170 GB?
Third-party pages quote anywhere from 82.5 GB to 170 GB for this model, and at least one of them covers V4.1 Flash, a different repo whose card was not read for this page. Those pages were not read here, so the reason for each figure is unknown. Arithmetic only shows what the low end would imply: 82.5e9 × 8 ÷ 284e9 = 2.32 bits per parameter, which is below the 4.40 bits per parameter average of the official 160 GB weights, so it cannot describe those weights. Treat only the primary numbers as sourced: 160 GB (Hugging Face repo) and 156.40 GB (LM Studio download).
Can I run DeepSeek V4 Flash on a DGX Spark?
Not the official weights on one unit. NVIDIA’s DGX Spark page lists 128 GB of coherent unified LPDDR5x at 273 GB/s and says one 128 GB unit runs models “up to 200B parameters”. V4 Flash is 284B. Arithmetic agrees: 128e9 × 8 ÷ 284e9 = 3.61 bits per parameter, below even a flat 4-bit 142 GB file, before cache. The same page lists two units (256 GB) as “up to 400B parameters”, which covers 284B; whether two linked units run V4 Flash well is not stated and speed is unknown. Compare the boxes on dgx-spark-vs-strix-halo.
What can I run on a $15k V4 Flash budget?
This page does not quote prices, so the useful question is what memory pool you need: at least about 162 GB for weights, plus cache. NVIDIA lists the RTX PRO 6000 at 96 GB GDDR7, 1792 GB/sec, 600 W. One card (96 GB) is short by about 66 GB. Two cards are 192 GB, which leaves 192 − 162 = 30 GB for KV cache on paper. Whether a split across two cards works in your runtime is untested here. With less GPU memory, the alternative is MoE offload: experts in system RAM, as on moe-cpu-offload; you then need about 160 GB of RAM (ram-vs-vram) and speed is unknown. Check current retailer price for any combination.
How much VRAM does DeepSeek R1 need?
The full R1 671B needs 671e9 × 8 ÷ 8 = 671 GB at 8 bits, and its repo is 689 GB. At Q4_K_M, using llama.cpp’s 4.8944 bits per weight (a figure the README gives for Llama 3.1 8B, so applying it to R1 is an approximation), the weights are 671e9 × 4.8944 ÷ 8 ÷ 1e9 = 410.5 GB. That is far beyond any single consumer or workstation GPU. R1 KV cache is not computed here; its config was not read, so it is unknown.
How much VRAM for each R1 distill (70B, 32B, 14B, 8B)?
The distill cards (Llama-70B, Qwen-32B, Qwen-14B, Llama-8B) give only rounded counts (71B, 33B, 15B, 8B) and BF16 weights. The table uses those rounded counts, so it overstates slightly. KV uses each config.json (80/64/48/32 layers, 8 KV heads, head dimension 128, 2 bytes per value; head dimension is stated for the 70B and computed as hidden size ÷ heads for the others).
| Distill | Q4_K_M | Q8_0 | BF16 | KV at 4,096 | KV at 32,768 | Q4_K_M at 4,096, with 1 to 2 GB overhead |
|---|---|---|---|---|---|---|
| Llama-70B (71B) | 43.4 GB | 75.4 GB | 142.0 GB | 1.34 GB | 10.74 GB | 45.8 to 46.8 GB |
| Qwen-32B (33B) | 20.2 GB | 35.1 GB | 66.0 GB | 1.07 GB | 8.59 GB | 22.3 to 23.3 GB |
| Qwen-14B (15B) | 9.2 GB | 15.9 GB | 30.0 GB | 0.81 GB | 6.44 GB | 11.0 to 12.0 GB |
| Llama-8B (8B) | 4.9 GB | 8.5 GB | 16.0 GB | 0.54 GB | 4.29 GB | 6.4 to 7.4 GB |
Bits per weight are 4.8944 (Q4_K_M) and 8.5008 (Q8_0) from the README. For the 70B, the README’s measured Llama 3.1 70B Q4_K_M file is 43.1 GB, close to the 43.4 GB estimate; the full walk-through is vram-for-70b-model. So the 70B needs a 48 GB class setup (48gb-vram-llm) or a split (multi-gpu-llm); the 32B at 22.3 to 23.3 GB lands within 1 GB of a 24 GB card, which this site does not call a fit (24gb-vram-llm, vram-for-32b-model). The 14B and 8B suit 16 GB and smaller cards. Distills are dense, so none of this offload math is MoE-specific.
Questions
How much VRAM does DeepSeek V4 Flash need?
About 160 GB for the weights, because the official Hugging Face repo is 160 GB (284B parameters, FP4 experts, FP8 elsewhere), or roughly 161 to 162 GB with the assumed 1 to 2 GB overhead. KV cache is unknown because the model compresses it per layer and the card gives no figure.
How much VRAM for DeepSeek R1 70B?
The distill’s rounded 71B parameters at Q4_K_M are 43.4 GB, and with 4,096 tokens of KV cache (1.34 GB) plus 1 to 2 GB of assumed overhead the total is 45.8 to 46.8 GB. At 32,768 tokens the cache alone is 10.74 GB. That needs 48 GB of memory or a multi-GPU split.
What GPU do I need to run DeepSeek R1?
It depends which R1. The full 671B needs about 410.5 GB even at Q4_K_M, which is multi-GPU or large-RAM territory. The 70B distill needs about 46 to 47 GB, the 32B about 22 to 23 GB, the 14B about 11 to 12 GB and the 8B about 6 to 7 GB at Q4_K_M and 4,096 context.
Can I run DeepSeek V4 Flash on a DGX Spark?
Not on one 128 GB unit with the official weights: 160 GB does not fit in 128 GB, and NVIDIA lists one unit for models up to 200B parameters against V4 Flash’s 284B. NVIDIA lists two units (256 GB) up to 400B parameters; speed on that setup is unknown.
What can I run on a $15k DeepSeek V4 Flash budget?
Prices are not quoted here, so check current retailer price. The target is a memory pool of at least about 162 GB for the weights plus cache room. Two 96 GB RTX PRO 6000 cards give 192 GB on paper, and CPU offload of the experts needs about 160 GB of system RAM; tokens per second are unknown for both.
Does V4 Flash activating only 13B parameters cut VRAM?
No. All 284B parameters must be stored somewhere; only the 13B used per token are read each step. That is why the memory need follows the 160 GB file while offload setups can trade VRAM for system RAM.
Sources read on 2026-10-13
- DeepSeek-V4-Flash model card, its config.json and file tree: 284B total, 13B active, FP4 + FP8, 1M context, MIT, 291B safetensors metadata, 160 GB across 46 shards, layers, experts, KV compression schedule.
- LM Studio deepseek-v4-flash: 156.40 GB download, “at least 156 GB of RAM”.
- Ollama deepseek-v4-flash: no sizes; retired August 27, 2026.
- DeepSeek-R1 card and file tree: 671B, 37B active, 128K, FP8, 689 GB, 163 shards, distill table.
- The four distill cards linked above and each repo’s
config.json: rounded parameter counts, BF16, layers, KV heads. - llama.cpp quantize README: bits per weight, Llama 3.1 70B Q4_K_M 43.1 GB.
- NVIDIA DGX Spark: 128 GB, 273 GB/s, model-size table.
- NVIDIA RTX PRO 6000: 96 GB GDDR7, 1792 GB/sec, 600 W.
- Not read: the V4.1 Flash and V4-Flash-0731 cards, and the secondary pages behind the 82.5 to 170 GB spread.
Related
- moe-cpu-offload explains keeping experts in system RAM.
- vram-for-70b-model and vram-for-32b-model go deep on two distill sizes.
- dgx-spark-vs-strix-halo compares the unified-memory boxes.
- 48gb-vram-llm and 24gb-vram-llm size by card.
- ram-vs-vram and multi-gpu-llm cover memory types and splits.
- new-models tracks releases.
- vram-requirements and llm-quantization-explained give the general formula; methodology is the citation rule.