Running DeepSeek V4 Flash 284B Locally – Performance Beats V4 Pro

Model specifications

Flash 0731 retains the same architecture as the preview: 284 B total parameters, 13 B active parameters, and a 1 M context window. The update mainly improves Agent capability, and benchmark results show Flash 0731 already exceeds the V4 Pro preview.

Quantization details

Flash uses quantization‑aware training; the official checkpoint stores routing experts (96 % of the model) natively as MXFP4, while the remaining parts use FP8 or BF16. Unsloth’s UD‑Q8_K_XL quantization reuses these MXFP4 expert weights without any additional quantization, making the conversion lossless— all 1 328 tensors match the official weights bit‑for‑bit, and inference KL divergence is approximately zero.

Hardware requirements for local execution

Unsloth provides several quantization levels with total memory (RAM + VRAM or unified memory) requirements:

1‑bit: 92 GB

2‑bit: 102 GB

3‑bit (UD‑IQ3_XXS): 110‑135 GB

4‑bit (UD‑Q4_K_XL, near‑lossless): 162 GB

Q8_K_XL (lossless): 169 GB

The 3‑bit version can run on a machine with 128 GB memory, while the 4‑bit lossless version requires 169 GB. Detailed usage guides for llama.cpp and Unsloth Studio are provided.

Performance surpasses V4 Pro

Official benchmark table shows Flash 0731 achieving 82.7 % on Terminal Bench 2.1, compared with 72.1 % for the V4 Pro preview—a 10‑point gain. DeepSWE improves from 12.8 % to 54.4 %, and NL2Repo rises from 38.5 % to 54.2 %. These gains stem mainly from the upgraded Agent capability rather than simple model scaling.

Native Codex support

V4 Flash natively supports OpenAI’s Responses API format and is fully compatible with Codex. A configuration script sets DeepSeek as the default Codex model with a single command.

# macOS / Linux
bash <(curl -fsSL https://cdn.deepseek.com/api-docs/codex-deepseek-setup-en.sh)

# Windows PowerShell
irm https://cdn.deepseek.com/api-docs/codex-deepseek-setup-en.ps1 | iex

The script automatically backs up existing configuration, writes the model directory and config.toml, and validates the syntax. Developers can also manage the model via tokenbank without manual steps.

Conclusion

GGUF files are available on Hugging Face at https://huggingface.co/unsloth/DeepSeek-V4-Flash-0731-GGUF.