readygpu

DeepSeek V4 Flash

290.9B total, 20.4B active · 1024k context · mit · released 2026-04-22

chatcodingagenticreasoning

What to run it on in India

Fully in memory at 8k context, priced at retailers today.

Something we price holds it fully in memory, but none reaches 15 tok/s, our bar for recommending it. The table below has the speeds.

All hardware that runs it

Everything we price, best value first. Two cards split the model by layers.

HardwareQuantFitsDecodePricetok/s per 1,000
Mac Studio M4 Max 128GBUD-Q2_K_XLfits28 tok/s––
Framework Desktop Ryzen AI Max+ 395 128GBUD-Q2_K_XLfits11 tok/s––
Mac Studio M3 Ultra 512GBUD-Q4_K_XLfits28 tok/s––
NVIDIA DGX Spark 128GBUD-Q2_K_XLfits14 tok/s––
GMKtec EVO-X2 Ryzen AI Max+ 395 128GBUD-Q2_K_XLfits11 tok/s––
Mac Studio M3 Ultra 256GBUD-Q4_K_XLfits28 tok/s––
NVIDIA RTX PRO 6000 BlackwellUD-Q2_K_XLfits89 tok/s––
ASUS Ascent GX10 128GBUD-Q2_K_XLfits14 tok/s––

32 other cards and machines we price cannot hold it at this context.

Memory needed

Weights, KV cache and runtime buffers at each context length (KV cache at fp16).

QuantFile@8k@32k@64k@128k
UD-Q4_K_XLgguf144.4 GB145.9 GB148.0 GB150.9 GB156.6 GB
UD-Q2_K_XLgguf90.2 GB91.6 GB93.8 GB96.6 GB102.4 GB
UD-Q3_K_Mgguf119.3 GB120.7 GB122.9 GB125.7 GB131.5 GB
mlx-4bitmlx141.1 GB142.5 GB144.7 GB147.5 GB153.3 GB

Run it

Copy a command for the runtime you use.

  • UD-Q4_K_XL
    • Ollamaollama run hf.co/unsloth/DeepSeek-V4-Flash-0731-GGUF:UD-Q4_K_XL
    • LM Studiolms get unsloth/DeepSeek-V4-Flash-0731-GGUF@ud-q4_k_xl
    • llama.cppllama-server -hf unsloth/DeepSeek-V4-Flash-0731-GGUF:UD-Q4_K_XL
  • UD-Q2_K_XL
    • Ollamaollama run hf.co/unsloth/DeepSeek-V4-Flash-0731-GGUF:UD-Q2_K_XL
    • LM Studiolms get unsloth/DeepSeek-V4-Flash-0731-GGUF@ud-q2_k_xl
    • llama.cppllama-server -hf unsloth/DeepSeek-V4-Flash-0731-GGUF:UD-Q2_K_XL
  • UD-Q3_K_M
    • Ollamaollama run hf.co/unsloth/DeepSeek-V4-Flash-0731-GGUF:UD-Q3_K_M
    • LM Studiolms get unsloth/DeepSeek-V4-Flash-0731-GGUF@ud-q3_k_m
    • llama.cppllama-server -hf unsloth/DeepSeek-V4-Flash-0731-GGUF:UD-Q3_K_M
  • mlx-4bit
    • LM Studiolms get mlx-community/DeepSeek-V4-Flash-4bit
    • MLXmlx_lm.server --model mlx-community/DeepSeek-V4-Flash-4bit

Notes

Estimates are approximate. Sparse or compressed attention: KV read per token is estimated as full attention. deepseek_v4 layout is new; verify against measured results.