DeepSeek V4 Flash 0731 / deployment worksheet

13B active. 155.43 GiB resident.

Q: How much GPU memory does DeepSeek V4 Flash need?

The official DeepSeek V4 Flash 0731 repository contains 48 safetensors files totaling 155.43 GiB. Two 80 GiB-class GPUs provide only 144 GiB when 90 percent of listed memory is usable, so three 80 GiB-class GPUs are the first practical all-GPU planning tier before measuring the exact runtime.

Q: Why does a 13B-active model need the full 155.43 GiB checkpoint?

DeepSeek V4 Flash is a mixture-of-experts model with 284 billion total parameters and about 13 billion activated per token. Routing lowers per-token compute, but the deployed checkpoint still contains the inactive experts, embeddings, attention layers, routing weights, and quantization metadata.

Q: How does the one-million-token context affect DeepSeek V4 Flash VRAM?

DeepSeek officially supports one million tokens and uses hybrid compressed attention. A portable calculator built from the exposed 43 layers, one KV head, and 512 head dimension produces an 86 GiB dense-cache comparison at one million tokens, but that is not a claim about DeepSeek's optimized HCA runtime allocation. Benchmark the exact engine and cache format.

Q: What changed in DeepSeek V4 Flash 0731?

DeepSeek says the 0731 checkpoint is the official release that supersedes the preview, substantially improves agentic capabilities, and attaches the DSpark speculative-decoding module. The official safetensors total is 155.43 GiB, 6.77 GiB larger than the 148.66 GiB preview artifact.

DeepSeek's official 0731 release supersedes the preview and adds DSpark speculative decoding. Size its 284B-total, 13B-active mixed FP4/FP8 checkpoint before cache and runtime reserve.

Open the editable 128K plan →

Quick answer / official artifact

Three 80 GiB-class GPUs are the first clean planning tier.

Two devices at 90% usable memory expose 144 GiB—less than the 155.43 GiB artifact itself. These editable profiles keep the exact 0731 checkpoint override and vary only the cached context and runtime reserve.

32K comparison · 3 × 80 GiB 173.93 GiB 155.43 checkpoint + 2.69 cache + 10% reserve → 128K comparison · 3 × 80 GiB 199.42 GiB 155.43 checkpoint + 10.75 cache + 20% reserve → 384K comparison · 4 × 80 GiB 225.22 GiB 155.43 checkpoint + 32.25 cache + 20% reserve → 1M comparison · 5 × 80 GiB 289.72 GiB 155.43 checkpoint + 86.00 cache + 20% reserve →

Claim vs deployment input

Active parameters describe compute. The files describe residency.

DeepSeek publishes the model shape and precision mix. The official repository exposes the actual artifact size. Runtime workspace, cache policy, tensor parallelism, quantized kernels, and offload determine what happens after the files load.

Swipe the matrix horizontally →

Input	Official value	What it changes	What it does not prove
Total parameters	284B	Checkpoint and resident-weight scale	Per-token compute
Activated parameters	13B	MoE work routed per token	Weight memory to load
Release status	0731 official; supersedes preview	Current checkpoint and agentic release	Hardware throughput
Released precision	FP4 experts + mostly FP8 elsewhere	Artifact size and supported kernels	Runtime overhead
Official checkpoint	155.43 GiB	Exact downloadable weight floor	Cache, workspace, or latency
Maximum context	1,048,576 tokens	Possible context boundary	Preallocated cache policy

The artifact total was calculated from the 48 safetensors file sizes in DeepSeek's official 0731 Hugging Face repository on August 7, 2026. Recheck the repository if DeepSeek revises the release.

Read the estimate

Active is not resident.
MoE routing reduces work per token without deleting inactive experts from the artifact.

Capacity is not speed.
CPU or unified-memory offload can make a model load while missing the latency target.

0731 replaces the preview.
The official release adds the attached DSpark speculative-decoding module.

Context comparison boundary

DeepSeek's hybrid attention is not a standard dense KV cache.

Official architecture
43 layers · 1 KV head · 512 head dimension

Official context
1,048,576 tokens · hybrid HCA + CSA

Portable comparison only
The editable calculator applies the standard dense-cache formula to the exposed dimensions so context profiles are comparable. It does not reproduce DeepSeek's compressed-attention runtime.

DeepSeek reports that V4-Pro's one-million-token KV cache is 10% of DeepSeek-V3.2 under its optimized architecture. That first-party relative claim is evidence that a generic dense-cache estimate and the production HCA allocation are different measurements.

Primary-source inputs

Use the exact artifact. Measure the exact runtime.

Release status, total and active parameters, precision mix, DSpark integration, context support, architecture, and inference instructions come from DeepSeek. File sizes come from the official 0731 model repository—not a third-party VRAM table.

Open the GPU memory calculator → Open the KV-cache calculator → Compare gpt-oss hardware requirements → Compare GLM-5.2 FP8 and H200 requirements → Official DeepSeek V4 Flash 0731 model card → Official DeepSeek V4 Flash 0731 configuration → Official 0731 checkpoint file inventory → Official local inference instructions → DeepSeek V4 technical report →

How much GPU memory does DeepSeek V4 Flash need?

The official 0731 mixed FP4/FP8 artifact is 155.43 GiB before cache and runtime workspace. At 90% usable memory, two 80 GiB-class GPUs expose only 144 GiB, so three are the first clean all-GPU planning tier.

Can DeepSeek V4 Flash run on two 80 GB GPUs?

Not as a straightforward full-checkpoint GPU-resident plan with sensible reserve. Offload can change whether the model loads, but then measured throughput and latency—not nominal VRAM alone—decide whether the deployment works.

Why does a 13B-active model need the full 155.43 GiB checkpoint?

Thirteen billion describes the experts activated for a token. The released artifact still contains 284 billion total parameters plus attention, routing, embeddings, the attached DSpark module, and quantization metadata.

How does the one-million-token context affect DeepSeek V4 Flash VRAM?

The official model supports one million tokens through hybrid compressed attention. The calculator profiles are transparent dense-cache comparisons, not measured HCA allocations; validate the exact runtime, cache format, and concurrency.

What changed in DeepSeek V4 Flash 0731?

DeepSeek says 0731 is the official release that supersedes the preview, substantially improves agentic capabilities, and attaches the DSpark speculative-decoding module. Its official artifact is 6.77 GiB larger than the preview.