by deepseek-ai

DeepSeek V4 Flash 0731

deepseek-ai/DeepSeek-V4-Flash-0731

All figures are estimates. Real memory use and speed depend on your drivers, settings and other programs running.

Parameters: 304B
Licence: mit
Published: Jul 31, 2026

The licence is the model maker’s. Check its terms before you use the model.

Memory needed by quantization

At 8,192 tokens of context, including 0.8 GB for the runtime.

Quantization	Model file	Total
Q4_K_M	184.4 GB	203.7 GB(rough)
Q5_K_M	216.3 GB	238.8 GB(rough)
Q6_K	249.4 GB	275.2 GB(rough)
Q8_0	323.2 GB	356.3 GB(rough)
F16	608.4 GB	670.0 GB(rough)

Hardware that runs it on the graphics card

At Q4_K_M and 8,192 tokens of context.

None of the hardware we list runs it fully on the graphics card at this setting.

Check it on your computer

The same model through an API

Cheapest seller: Merge Gateway, $0.42 a month at the default use (200 requests a day).

See the prices on the Model Price Map

Local Setup Kit

Scripts for your system that check your runtime, download this model from Hugging Face, verify the file, run it on this computer only, and remove it again.

Install, run, speed-test and uninstall scripts for the system and runtime you choose.
Settings for the quantization and context length you choose.
A step-by-step README in your language.
A safety report: what each script was checked for before it reached you.
The model’s licence name and link. The model itself downloads from Hugging Face.

SystemRuntimeQuantizationContext length

One-off payment of $4.99. Download right after payment, as often as you need for 7 days (up to 10 downloads).

If we cannot build your kit, we refund it in full. Other refund requests within 14 days are decided under our refund policy.

I want the kit now and agree that delivery starts immediately after payment. I understand that I then lose my right to withdraw from the purchase.

Secure payments by Stripe. Prices are shown in your local currency.

Model details come from Hugging Face’s public listings (names, sizes, licences). Every model page links to the original.

What's new

2026-10-13You can now share a link that opens in the language you chose: switching languages updates the page address.