DeepSeek's 'Flash' Model Needs 475GB Before Anyone Runs It

DeepSeek says its new V4.1-Flash model activates just 8 billion parameters to read a prompt and 16 billion to write a response. Its total footprint, once every routed expert and the model's separate long-term memory module are counted, is roughly 763 billion parameters. The gap between those two numbers is why "Flash" means something different on DeepSeek's own servers than it does to anyone trying to run the model themselves.

DeepSeek announced V4.1-Flash on Sept. 10, describing a 552-billion-parameter mixture-of-experts model built on what it calls a Causal Encoder-Decoder architecture: 8 billion active parameters process each input token, 16 billion generate each output token. The model card DeepSeek published on Hugging Face adds a separate 196-billion-parameter component it calls Engram, described as conditional long-term memory, bringing the full repository to roughly 763 billion parameters under an MIT license. Native context runs to 1 million tokens. Engram itself traces to a January 2026 paper co-authored by DeepSeek founder and chief executive Wenfeng Liang, which described the technique as a lookup-based complement to mixture-of-experts routing and reported gains of 3.4 points on the MMLU benchmark and 5.0 points on BBH in earlier, smaller-scale testing. DeepSeek's own benchmark comparisons, which The Terminal could not independently verify, show V4.1-Flash beating the outgoing V4-Pro on cost, speed and total runtime, a claim that would matter more to potential buyers if any of them could load the smaller model on their own hardware.

890 bytes a token is real, but it isn't the whole bill

DeepSeek's headline efficiency claim is a global key-value cache of 890 bytes per token, which the Hugging Face card says is close to a quarter of what the prior V4-Flash needed. DeepSeek's own announcement puts the reduction at a quarter of the HBM and an eighth of the SSD spillover storage compared with that predecessor. At 1 million tokens of context, that works out to under 1 gigabyte of cache. That number describes the cost of holding a conversation in memory once the model is already loaded. It says nothing about the cost of loading the model in the first place, which is the number that decides whether the cache savings reach anyone outside DeepSeek's own data centers.

What it actually takes: 475 gibibytes, minimum

Modem Guides put a figure on that cost on Sept. 10: the model as shipped, in mixed FP8 and FP4 precision, needs about 475 gibibytes of memory, and the Engram tables alone need roughly 385 gibibytes even quantized down to FP4. That rules out every machine marketed as a run-large-models-at-home box. An Apple M5 Max with 128 gigabytes of unified memory is not close. An M5 Ultra with 512 gigabytes technically has the capacity, Modem Guides found, but no software yet exists to load the architecture on it. A Ryzen AI Max+ 395 mini PC, the enthusiast favorite for squeezing large open-weight models onto a desk, tops out at 128 gigabytes and hits the same wall.

Hardware	Memory	Fits V4.1-Flash?
Apple M5 Max	128 GB unified	No
Ryzen AI Max+ 395 mini PC	128 GB unified	No
Apple M5 Ultra	512 GB unified	Capacity yes, no runtime
DeepSeek reference stack	4-8 datacenter accelerators	Yes, via SGLang preview

1. Model as shipped (FP8/FP4)475 GB
2. Engram tables alone (FP4)385 GB
3. Apple M5 Ultra ceiling512 GB
4. Apple M5 Max ceiling128 GB
5. Ryzen AI Max+ 395 ceiling128 GB

Source: Modem Guides, accessed 2026-09-12

Active parameters set speed. Total parameters set the memory bill.

Modem Guides put the tension precisely: "active parameters set speed. Total parameters set the memory bill." DeepSeek's own reference implementation reflects that gap. It runs on eight-way tensor parallelism across four to eight datacenter accelerators through a preview build of SGLang, according to Modem Guides, hardware that costs more in a single node than most self-hosters spend on years of cloud inference. As of publication, no released inference engine supports the architecture DeepSeek calls deepseek_v41 at all. The closest precedent is V4-Flash's own predecessor: after its April 2026 release, mainline llama.cpp support took 66 days to land, Modem Guides found, on a model with a smaller memory footprint than V4.1-Flash's.

The timing compounds the problem. DeepSeek is retiring V4-Pro outright: starting Sept. 14, every request aimed at that model routes automatically to V4.1-Flash at V4.1-Flash's lower price, whether or not the requester has any way to run the new architecture themselves. For DeepSeek's own hosted API, none of this matters; the company controls the accelerators and the serving stack. For Qwen3.8-27B, the comparable open-weight release this month that at least fits on a single 24-gigabyte consumer GPU at 4-bit, the contrast is direct: fitting on hardware people actually own is a design choice DeepSeek didn't make this time.

What would change this read is a shipped inference engine, from DeepSeek or anyone else, that gets the 763-billion-parameter repository onto hardware smaller than a rack, the way cold-start caching work elsewhere in the industry has tried to shrink the gap between a model's size and its readiness to serve. Until an engine like that exists, "Flash" describes what happens on DeepSeek's servers, not what happens on anyone else's.