DeepSeek V4 Flash 0731
304.2B total, 33.6B active · 1024k context · MIT License · released 2026-07-31
chatcodingagenticreasoning
What to run it on in India
Fully in memory at 8k context, priced at retailers today.
Nothing we price holds this model fully in memory.
All hardware that runs it
Everything we price, best value first. Two cards split the model by layers.
Nothing we price can run it at this context.
Memory needed
Weights, KV cache and runtime buffers at each context length (KV cache at fp16).
| Quant | File |
|---|
Run it
Copy a command for the runtime you use.
No published builds to run yet.
Notes
The card ships no Jinja chat template, so prompts must be encoded with the Python scripts in the provided `encoding` folder. It includes a DSpark speculative decoding module; the example vLLM setup uses a 4×GB300 node, and the card recommends a maximum output length of 384K tokens at the high and max reasoning effort levels. Sparse or compressed attention: KV read per token is estimated as full attention. deepseek_v4 layout is new; verify against measured results.
Questions
- What licence is DeepSeek V4 Flash 0731 released under?
- MIT License, which allows commercial use.