• 3 % discount on prepayment
  • Insured shipping from Switzerland
  • Sustainable: refurbished instead of new
  • Warranty on all cards
← Back to overview

DeepSeek-V4-Flash-0731

DeepSeek
MIT

Open language model for chat and text. 3 sizes from 304.2B to 1.7T parameters, partly mixture-of-experts (only a part of the parameters is active per token).

  • Mixture-of-experts: 304.2B parameters in total, of which 33.6B active per token. Memory requirement depends on all parameters, speed on the active ones.

Settings

Quality
4 bit (Q4_K_M) is the most common choice: good quality with low memory requirements.
Context length: 8k tokens
4k256k
How much text the model keeps in view at once. More context needs more memory.
Recommendation
4× H100 80 GB

The model needs about 184.7 GB of memory. Hopper, FP8 support.

Weights
182.5 GB
KV cache
0.7 GB
Speed
~40.2 Tok./s

More suitable configurations

CardQuantityTok./sStatus
Tesla V100 32 GB8×15.4only 6 in stock
AMD Instinct MI50 32 GB8×11.8only 2 in stock
AMD Instinct MI60 32 GB8×10.92-3 weeks delivery time
AMD Instinct MI100 32 GB8×13.22-3 weeks delivery time
A40 48 GB8×14.82-3 weeks delivery time
RTX A6000 48 GB8×16.62-3 weeks delivery time
Quadro RTX 8000 48 GB8×12.22-3 weeks delivery time
L40 48 GB8×18.12-3 weeks delivery time
A100 40 GB8×18.32-3 weeks delivery time
A100 80 GB4×28.62-3 weeks delivery time
H100 80 GB4×40.2only 1 in stock