Download DeepSeek V4 Flash from HuggingFace & Run It Locally

DeepSeek hosts its open weights in one place: the official HuggingFace organization at huggingface.co/deepseek-ai. For the July 31, 2026 final release, the repository you want is deepseek-ai/DeepSeek-V4-Flash-0731 - the deepseek v4 flash huggingface release that went live under the MIT license with ~167 GB of weights across 48 Safetensors shards.^[1]

DeepSeek also maintains a model collection at huggingface.co/collections/deepseek-ai/deepseek-v4. As of August 1, 2026, the collection lists seven models spanning the Flash and Pro lines, including base variants for research and the fused DSpark checkpoints.^[2]

• deepseek-ai/DeepSeek-V4-Flash-0731 - the official final Flash release (July 31, 2026); the default choice for local inference
• deepseek-ai/DeepSeek-V4-Flash - the April 24 preview weights
• deepseek-ai/DeepSeek-V4-Flash-Base - base weights for research and fine-tuning (shows 292B params, FP8)
• deepseek-ai/DeepSeek-V4-Flash-DSpark - preview weights plus the speculative decoding draft module
• deepseek-ai/DeepSeek-V4-Pro / DeepSeek-V4-Pro-Base - the Pro line, still preview as of August 1, 2026
• deepseek-ai/DeepSeek-V4-Pro-DSpark - Pro preview weights plus draft module

Open the DeepSeek-V4-Flash-0731 model card and HuggingFace shows 304B params, ~167 GB, and 48 Safetensors shards. The April preview listed 284B. Nothing got bigger: the extra parameter count comes from the DSpark speculative decoding draft module bundled into the same repository.

Under the hood, DeepSeek V4 Flash is a 284B-total / 13B-active Mixture-of-Experts model with a native 1M token context window and up to 384K tokens of output. The 0731 release is architecturally identical to the April 24 preview. DeepSeek only re-post-trained the weights and attached the DSpark draft module, which shares the same structure as the DeepSeek-V4-Flash-DSpark repository.

The weights ship in mixed precision: FP4 for the MoE expert weights and FP8 for everything else. This is a natively quantized distribution, not a BF16 release. DeepSeek has never published a BF16 version of V4 Flash.

All 284B weights must be resident in memory even though only 13B activate per token. 'Flash' describes inference cost, not footprint. Plan for roughly 170-175 GB of total VRAM for the full-precision weights plus KV cache.

The realistic reference rig for the official FP4+FP8 weights is a 4x H200 node running vLLM with MoE expert parallelism.^[3] Two H200s (282 GB) or two RTX Pro 6000 Blackwell cards (192 GB) also work. A single 24 GB consumer card cannot run the full model.

If you want a smaller footprint, the GGUF route is the practical answer. Unsloth repackages the official quantization-aware-trained weights bit-exact at the expert level,^[4] so you do not lose precision where it matters most.

The official tool is huggingface-cli from the huggingface_hub package. Enable hf_transfer to parallelize downloads on fast connections, then point --local-dir at a target folder. The full 0731 snapshot is ~167 GB, so plan disk space and a stable network.

The equivalent Python call is snapshot_download(repo_id="deepseek-ai/DeepSeek-V4-Flash-0731", local_dir="./DeepSeek-V4-Flash-0731") from huggingface_hub.

In mainland China, use the ModelScope mirror instead for faster transfer: modelscope.cn/models/deepseek-ai/DeepSeek-V4-Flash. Weights on the official deepseek v4 flash huggingface channel and the ModelScope path are the same files.

pip install huggingface_hub hf_transfer
export HF_HUB_ENABLE_HF_TRANSFER=1   # speeds up downloads on 1Gbit+ connections

# Full official 0731 release (~167 GB, includes DSpark draft module)
huggingface-cli download deepseek-ai/DeepSeek-V4-Flash-0731 --local-dir ./DeepSeek-V4-Flash-0731

# Weights-only download (skip non-weight files)
huggingface-cli download deepseek-ai/DeepSeek-V4-Flash-0731 \
  --include="*.safetensors" "*.json" "*.py" --local-dir ./deepseek-v4-flash-0731

The official model card publishes a reference vLLM command built for a single-node 4x GB300 setup. The critical piece for the 0731 weights is the DSpark speculative decoding flag - one line enables draft-based generation.

DSpark is a single flag: --speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'. vLLM 0.25.0 and newer support DSpark drafting. ROCm targets (MI325X / MI355X) require vLLM 0.26.0.

vLLM exposes the model through an OpenAI-compatible interface when you pass --tokenizer-mode deepseek_v4, so standard chat completions calls just work. Add --reasoning-parser deepseek_v4 and --tool-call-parser deepseek_v4 for reasoning and tool calls.

vllm serve deepseek-ai/DeepSeek-V4-Flash-0731 \
  --trust-remote-code --kv-cache-dtype fp8 --block-size 256 \
  --data-parallel-size 4 --enable-expert-parallel \
  --moe-backend deep_gemm_mega_moe \
  --attention-config '{"use_fp4_indexer_cache": true}' \
  --speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'

If 167 GB of FP4+FP8 weights is too much, Unsloth publishes repackaged GGUF quants of the 0731 release at unsloth/DeepSeek-V4-Flash-0731-GGUF.^[5] Because the official weights are quantization-aware-trained, the routed experts (about 96% of the model) are natively MXFP4 - Unsloth keeps them bit-exact while rounding only the non-expert remainder.

The UD-Q8_K_XL quant is 161.9 GB and 100% bit-exact to the official weights (KL divergence ≈ 0). UD-Q4_K_XL is 155.1 GB near-lossless, and UD-IQ3_XXS is 103 GB with a recommended 110 GB of memory.

Beware community quants that down-convert the experts themselves: re-quantizing routed experts to Q4_K or IQ2_XXS rounds 5% and over 30% of the weights respectively. Stick to quants that preserve expert bit-exactness when quality matters.

hf download unsloth/DeepSeek-V4-Flash-0731-GGUF --local-dir ./deepseek-v4-flash-gguf --include "*UD-Q8_K_XL*"

# llama.cpp: control thinking / reasoning effort
--chat-template-kwargs '{"enable_thinking":false}'   # disable thinking
--chat-template-kwargs '{"reasoning_effort":"max"}'  # set reasoning tier

Downloading 167 GB only makes sense if you have the hardware or a real use case. For most developers, the official API is the pragmatic choice: model id deepseek-v4-flash, input at $0.14 per 1M tokens, output at $0.28 per 1M tokens, and a ~98% cache-hit discount that drops input to $0.0028.^[6]

The same model id deepseek-v4-flash powers both the 0731 and preview builds on the API; DeepSeek upgraded the weights without renaming the endpoint, so existing calls are unaffected. Note that the legacy ids deepseek-chat and deepseek-reasoner retired on July 24, 2026 at 15:59 UTC.

If you only have a 24-48 GB consumer card, skip the full weights. Use the API or a smaller distill model instead - a full 284B model will not fit. Independent testing by Artificial Analysis puts the 0731 release at an Intelligence Index of 50, up 10 points from the April preview; that is the model you get on either path.