Can You Self-Host DeepSeek V4-Flash? It Needs About 200GB

Four days ago this site argued that open weights at the frontier do not buy you the ability to run the model. Kimi K3 had just landed with 2.8 trillion parameters and needed about 1.4 terabytes of fast memory before it could read a token, which works out to roughly eighteen of the 80GB cards data centers run on.

Open, in that case, bought provider choice and data control. It did not buy a server in your closet.

That argument was about a constraint, and constraints have numbers. Here is the other number: 200 gigabytes.

That is the size of the DeepSeek V4-Flash download, listed for the July 31 build by vLLM, the standard open-source tool teams use to host models like these.^[3] Same month, same category of thing, roughly a seventh of the memory.

What actually shipped on July 31?

An API, not a model. DeepSeek put a retrained build of V4-Flash into public beta, matched it to the request format OpenAI's tools expect, and tuned it for coding-agent work. The weights themselves were already public, and the retraining did not change the size of the thing.

That detail is easy to skip past and it is the most interesting part. V4-Flash shipped on April 24 with the V4 preview: weights free to download from Hugging Face, the public library for AI models, under the permissive MIT license.^[1][5]

The small near-frontier open model was not a July surprise. It had been sitting there for three months while the conversation was about something else.

Why does the size of the weights decide who can host a model?

Because a model has to sit fully in fast memory before it answers anything, and the weights set the floor. V4-Flash is a mixture-of-experts design: a roster of specialists where each word wakes only a few (13 billion of its 284 billion parameters). The whole roster still loads into memory. Speed follows the active few. Hostability follows the total.

That is the number the July releases moved. The download ships heavily compressed (most of it stored at a quarter of the usual precision), which lands the July 31 build at about 200GB.^[3] vLLM's list of tested setups runs from eight-card server racks down to a single workstation card with 252GB of memory.

Compare that to a server built around eight 192GB cards, and the difference is not a discount. It is a different class of buyer.

None of this makes running it at home a solved problem. A 252GB card is not a laptop, and compression work aimed at smaller machines is a direction, not a promise. What changed is the number of parties who can host the thing at all: a shorter list than "everyone," a much longer one than "the biggest cloud providers."

Is the model actually good enough to matter?

Close enough that the question is live, which is new. DeepSeek reports 79.0 on SWE-bench Verified, the most-cited test of real-world coding, and 91.6 on LiveCodeBench, another coding benchmark.^[5] Artificial Analysis, an independent evaluator, scores the July 31 build at 50 on its intelligence index and ranks it third of 101 models in its class.^[4]

The pricing is doing its own arguing. Fourteen cents per million input tokens and 28 cents per million output, against medians of 43 cents and $1.20 for models Artificial Analysis groups it with.^[4]

Developer reaction has been about the size rather than the scores, which is the part worth noticing.

Those comparisons are reactions to published numbers, not controlled evaluations, and the coding figures above are the lab's own. Early enthusiasm outruns sustained use fairly reliably. The defensible claim is narrower and still significant: a model in this size class is now in the conversation with models many times larger, and that was not true this spring.

Is a price cut the same as an efficiency gain?

No, and the two showed up 24 hours apart. OpenAI cut prices on two of its models by up to 80% on July 30, and we wrote about why that cut misses most agent work. A price is a decision a vendor makes about margin, and a decision can be unmade on a Tuesday.

An efficiency gain is a different kind of event. It lowers what the work costs to run anywhere, it ships inside weights anyone can download and host, and no pricing page governs it. When today's prices are held down by capital rather than costs, that distinction is the whole game.

Trade coverage has tied part of the OpenAI cut to competitive pressure from Chinese labs. Whether or not that holds, the sequence is worth sitting with: a cheaper invoice one day, a smaller model the next.

What this changes for a team

Not much this week, and quite a lot over a year. The move is not to switch models. It is to make sure switching stays possible, which is a different kind of work and mostly happens before you need it.

Three things worth doing while this is fresh:

• Find out what your work is actually pinned to. Usually it is one vendor's request format and the way it wires into your tools, not the model.
• Price the same task across two or three providers hosting the same open weights. That number is your real leverage.
• Watch the size trend rather than any single release. The direction is what matters, and portability is a posture you build in advance.

Our July 27 argument still holds for models at the very top: at 2.8 trillion parameters, open buys you provider choice, not independence. What we underweighted is that the frontier is not one point. It is a band, and the lower edge of that band is now arriving in packages small enough that a lot of people can hold them.

The open-weights conversation has spent a year on licenses and a month on price. The variable that decides whether open means usable was always size.