DeepSeek V4.1-Flash hardware requirements

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 The first question everyone asked was whether it would run on a 128GB machine, because the last model called Flash did.

It will not. V4.1-Flash ships as 510GB of FP8 weights across 48 files, and the compression step has already been spent.² But it does run locally today, on hardware you can buy: a community 2-bit build has been measured at 9.46 tokens per second on a 256GB Mac Studio M3 Ultra.⁵

• 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 what it costs per token on an API. They do not set the memory bill.

There is a second, less obvious reason the file is so large. The model carries a two-layer engram, a hashed n-gram embedding whose lookup tables alone account for roughly 196B parameters — about forty percent of the checkpoint.⁵ That is memory you hold and rarely compute with, and it is why the download is bigger than the backbone count suggests.

The previous model made the trap easy to fall into. DeepSeek-V4-Flash-0731 was 284B total with 13B active, and ran losslessly in 169GB.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. DeepSeek did that before you downloaded it: the repository ships in FP8 with the expert layers already compressed.² There is less left to take out than the usual four-to-one cut people expect from a 4-bit GGUF, which is why the community builds below do not fall as far as you would guess.

What actually runs it

This is the part 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 The deepseek_v41 architecture exists in no mainstream runtime — not Transformers, not mlx-lm, not mlx-vlm, and not llama.cpp.⁴

That rules out the easy routes. Ollama, LM Studio and anything else built on llama.cpp cannot load it, and there is no usable GGUF. It does not rule out running it.

On Apple silicon, the PipeNetwork port supplies the missing architecture and community MLX conversions supply the weights.⁴ On NVIDIA and AMD servers, SGLang publishes development container images for it — lmsysorg/sglang:dev-dsv41 and dev-dsv41-mi35x — ahead of a tagged release. vLLM has open pull requests and, as of 10 September, none merged.⁶

So this is early software on settled hardware. Expect to clone a repository rather than type one install command, and expect that to get easier over the next few weeks rather than harder.

Which machine

The 2-bit build is the one to start with, and it is genuinely good news. It is 238.8GB on disk and peaked at 166.69 GiB of process memory in testing, which puts it inside a Mac Studio M3 Ultra 256GB with room to spare. Its author measured 9.46 tokens per second with greedy decoding on exactly that machine.⁵ That is not fast, and for a 552B model answering from a desk it is remarkable.

The mixed 4/8-bit build asks for 427GB resident, which a Mac Studio M3 Ultra 512GB holds, tightly. Tight matters here: leave headroom for the context, because a long prompt at this context length is not free. The engram-6 build wants roughly 477GB and a terabyte-class machine, since a lazy forward pass can transiently need about twice the build size.⁴

Above 512GB in one box there are two honest routes. Thunderbolt 5 clustering pools unified memory across several Macs — macOS 26.2 added RDMA over Thunderbolt 5 and exo 1.0 supports it, so linked Mac Studio M3 Ultra 512GB machines add up.⁷ Or you build one workstation with the capacity: a Neox Metal on the Threadripper PRO platform takes eight channels of registered ECC memory, which is how you reach a terabyte in a single machine, and the same platform holds the RTX PRO 6000 cards that the SGLang route wants.

Settings

DeepSeek recommend a temperature of 1.0 with top-p at 0.95 or 1.0, and at least 256K of max_tokens.¹ Reasoning effort is a continuously controllable integer from 1 to 100 rather than the low/high/max switch other models use, so it is worth tuning rather than leaving at a default.

One number is unusually good. V4.1-Flash uses 890 bytes per token of global KV cache, which keeps a full million-token context under a gigabyte.¹ On most models of this size the context is what forces the next memory upgrade. Here it is the weights, and only the weights.

For the arithmetic behind all of this, see 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.