DeepSeek V4.1-Flash needs 510GB, and nothing can run it yet

DeepSeek released V4.1-Flash on 10 September 2026 under the MIT licence, with a one-million-token context window and native image understanding.¹ Within a day the questions started: will it run on a 128GB machine, will it run on a 5090, how small does it quantise.

The answer is that V4.1-Flash is a 510GB model. It ships as 48 files of FP8 weights, the compression step has already been spent, and as of today no released inference engine can load it at all.²⁴ This is worth understanding before you buy anything, because the name is doing a lot of work.

• Parameters552B8B in, 16B out
• Context1Mtokens
• LicenceMITno revenue trigger
• Weights510GBFP8, 48 files

Flash describes the speed, not the size

V4.1-Flash holds a 552B-parameter backbone and activates 8B parameters while it reads your prompt and 16B while it writes the reply.¹ Those small active figures are what "Flash" refers to. They set how fast it generates and how much it costs per token on an API.

They do not set the memory bill. A mixture-of-experts model chooses different experts for every token, so all of them must stay loaded. Active parameters set the speed; total parameters set the capacity you have to buy.

The previous model made this easy to miss. DeepSeek-V4-Flash-0731 was 284B total with 13B active.³ V4.1-Flash carries roughly double the backbone under the same product name.

The quantisation has already been spent

Normally a 552B model would be an invitation to quantise. Store each weight in four bits instead of sixteen and the footprint falls by three quarters.

DeepSeek did that before you downloaded it. The repository ships in FP8, and the expert layers are already compressed.² There is far less left to take out. One published analysis puts a further FP4 conversion at roughly 385GiB in total, against 510GB as shipped — a useful saving, and nothing like the four-to-one cut people expect from a 4-bit GGUF.⁴

Nothing runs it yet

This is the part that no specification sheet tells you. A new model needs its architecture written into whichever engine you use, and V4.1-Flash introduced a new one: a causal encoder-decoder, forty layers arranged as a twenty-layer causal encoder followed by a twenty-layer decoder.¹

Mainline llama.cpp has no deepseek_v41 architecture, and no open pull request adding one. Support for the older DeepSeek V4 series was a substantial piece of work in its own right, merged as PR #24162 to handle the compressed attention the series introduced.⁵ The GGUF repositories that have appeared for V4.1-Flash contain a README and no weights.⁴

So the honest position in September 2026 is that V4.1-Flash is a model you can serve on a GPU cluster with a preview build, or use through DeepSeek's API, and not yet one you can download and run on a workstation. That will change. When it does, the capacity it needs will not.

The DeepSeek you can run today

DeepSeek-V4-Flash-0731 is the one to actually put on a machine. It is 284B total with 13B active, MIT-licensed, and it has mature support.³

Its best property is unusual. The lossless 8-bit build needs 169GB, and the 4-bit build needs 162GB.³ Seven gigabytes separate a compromise from no compromise at all. There is almost no reason to run this model quantised. If you have the capacity for 4-bit, you have the capacity to run it exactly as DeepSeek shipped it, with the accuracy question removed from the table.

DeepSeek recommend a temperature of 1.0 and top-p of 1.0, or 0.95 for agentic work, with min-p at 0.01.³ If you use the Think Max reasoning mode, give it at least 384K of context to work in — the model will use it. Their DSpark speculative decoding roughly doubles decode speed and wants about 10GB of headroom on top.

What to buy for it

For V4-Flash-0731 at lossless Q8, you need 169GB in one addressable pool. A Mac Studio M3 Ultra with 256GB clears it with room for a long context. So does a Neox Metal with 256GB of ECC RDIMM and a card holding the attention layers, which is the cheaper route to the same capacity and the slower one, because the expert weights then read at the speed of your system memory rather than your VRAM.

For V4.1-Flash, the target is 512GB. That is the figure to plan against whether the eventual build is FP8 at 510GB or FP4 nearer 385GB, because you want the context and the working state above the weights, not squeezed in beside them. Today a 512GB machine means M3 Ultra, in stock; Apple's M5 Ultra is faster but its 512GB configuration is not due until late October.⁶ On the workstation side it means 512GB of registered ECC memory on an eight-channel platform.

The wider arithmetic behind all of this is in how much memory you need to run an open model locally. For the model currently at the top of the open leaderboards, see GLM-5.3 hardware requirements.