What is DeepSeek-V4.1-Flash, and why does a smaller memory cache make it cheaper?

DeepSeek released V4.1-Flash yesterday, and it comes with an unusual message for such a release. The company says this smaller model beats its own flagship, V4-Pro, on performance, cost, speed and total runtime in tests by several parties, so it is phasing V4-Pro out. Starting September 14, requests to deepseek-v4-pro go to the new model and are billed at its rates, until a V4.1-Pro launches.

The weights are on Hugging Face under the MIT licence. Weights are the learned numbers of a model, and MIT is one of the most permissive licences around.

What was released

A parameter is one learned number, and a token is a small piece of text. This is a mixture-of-experts model: 384 routed experts per layer, 6 of them chosen per token, plus one shared. The model also holds a separate 196-billion-parameter memory called Engram, consulted through lookups by token, which helps explain why Hugging Face's parameter count comes out higher than 552 billion. The download is about 510 GB. V4.1-Flash reads text and images natively, and it was pre-trained on 45 trillion tokens.

What a KV cache is, and why it costs money

To follow this release, you need one idea. When a language model reads your input, it builds stored notes on every token, called the KV cache (key-value cache). It consults them while writing. Long conversations and agent sessions hold huge caches, and that memory is costly: it sits on expensive high-bandwidth chips, or on fast disks when offloaded.

DeepSeek's changes target that cost directly. The company reports a global KV cache of 890 bytes per token. Here's a worked example of what that means for a long session:

Against the previous generation, DeepSeek says the cache needs a quarter of the high-bandwidth memory and an eighth of the disk storage. The card describes a roughly fourfold cut from V4-Flash and a 437-fold cut from DeepSeek-V1. That is a server-wide saving. A busy provider holding many sessions gains the most from it, and DeepSeek says those charges for cached input often make up a large part of agent bills.

How the design works

Three ideas do most of the work, in plain terms.

An encoder and a decoder that differ. The model's 40 layers split into a 20-layer encoder and a 20-layer decoder, and the stored notes come from the encoder's final stage instead of from every decoder layer. That is why reading activates 8 billion parameters and writing 16 billion. Agent work involves far more reading than writing, so a cheap reading path pays off.

Sharing notes across layers. A second version of its compressed sparse attention lets layers reuse each other's cached data and search results, instead of each keeping its own.

Drafting ahead. DSpark is speculative decoding, in which a quick draft proposes text and a check confirms it, to save time.

How strong is it?

DeepSeek's card includes comparisons of its base models on older benchmarks, and the V4.1-Flash base leads on BigCodeBench (60.6) and HumanEval (79.4), but V4-Pro's base keeps a clear lead on knowledge tests such as SimpleQA-Verified (55.2 against 42.3). So "ahead of Pro" holds on the agent results DeepSeek highlights more than on stored knowledge. The numbers are its own, and outside testers will need time to check them.

Effort can be set on a scale from 1 to 100, so you can pay for as much reasoning as a task needs.

What it takes to self-host

Set your expectations with the file size first. At about 510 GB, the weights need at least seven 80 GB graphics cards before any cache is counted, and DeepSeek's sample conversion uses an eight-way split across GPUs. DeepSeek's own reference code is described as readable, not a production engine, and the company says it is working with the open-source community on inference support. That means you should expect some rough edges at first if you serve it yourself.

The MIT licence itself makes no demands beyond keeping the notice. There are no revenue limits, attribution banners or model-hosting conditions, so a provider can host it freely. The training data isn't published, so this is open weight and not open source in the strict sense.

What does this release mean if you already use DeepSeek-V4?

If you call the API, check your settings now. The new model name is deepseek-flash. V4-Flash and the vision test version are retired, with their old names temporarily routed to V4.1-Flash. Because deepseek-v4-pro also reroutes on September 14, anyone who picked Pro for a reason should run their tests now. Scores on knowledge-heavy tasks may fall even if coding improves.

If you are self-hosting V4-Pro, nothing changes on its own, because the files you have keep working. The release does make it worth comparing with V4.1-Flash, which needs far less memory.

And the usual caution holds for the hosted service. Its data goes to DeepSeek under the terms covered in our earlier guide to trying DeepSeek-V4. The open weights give Canadian organizations another route, with a host or hardware they choose.