What hardware do you need to run DeepSeek V4-Flash?
DeepSeek V4-Flash at UD-Q4_K_M is 155 GB of weights, with a context ceiling of 1024k tokens. Not yet rated: released after our last ratings pass. The DeepSeek that replaced the R1 distills. Its compressed cache makes a million-token context practical.
How good is it, really?
On the Artificial Analysis Intelligence Index v4.3 it scores 35 (max effort; the 0731 build), which puts it in the Sonnet-class band. In the same league as the labs' mainstream models on this index. Score source. See the whole table.
- Summarising — not rated
- Translation — not rated
- Everyday coding — not rated
- Reasoning & maths — not rated
- Agentic work — not rated
What it costs either way
Renting the same model costs $0.065 per million input tokens and $0.18 per million output (OpenRouter, cheapest active endpoint, checked 2026-09-09). Buying a machine only beats that if you use it hard enough, for long enough, that the hardware price divides down below the rental bill.
Machines that run it
| Machine | Price | Speed at 32k | Pay-back | |
|---|---|---|---|---|
| Mac Studio M5 Ultra, 256GB | $10,799 | 49 tok/s estimated | Pays back in 1,056 years | Run the numbers |
One machine per family, cheapest first. Speeds are measured where a public benchmark exists and estimated from memory bandwidth otherwise; the calculator says which for any configuration.
The specifics
- Parameters
- 284B, of which 13B are active per token
- Quantisation
- UD-Q4_K_M
- Weights on disk
- 155 GB
- KV cache
- 0.2 GB at 32k context — Every layer compresses its cache by a different ratio: 21 layers keep one entry per four tokens, 20 keep one per 128, and two are a fixed 128-entry window. Summed, that is about 6.9 KB a token for the whole model — roughly a thirteenth of what the head counts imply. Read from the repo's own inference code.
- Maximum context
- 1024k tokens (1M)
- Licence
- MIT
- Sources
- source 1