Active experts change compute, not total storage

The report describes a large total parameter count with a smaller active portion during inference. Active compute and the memory needed to hold the model are different requirements.

A small active count does not make the full checkpoint small enough for any laptop.

Why the endpoint and checkpoint must be identified

A hosted service may modify prompts, precision, or serving settings. Compare the exact checkpoint and endpoint, rather than treating every service with DeepSeek in its name as equivalent.

What the report is about

The paper describes a large mixture-of-experts language model and the systems choices around training it, including multi-head latent attention and multi-token prediction. Those are design claims you can check against the paper, not against a screenshot of a chat.

Mixture-of-experts means the model does not run every parameter for every token. A parameter count and a per-token compute count are different numbers.

Quote the one the report actually uses.

Open weights, when the publisher releases them, come with a license. The architecture paper does not settle the license, the hosting provider’s retention, or whether a particular API is serving that file.

Match the checkpoint hash or the model id before you attribute a behavior to V3.

What not to borrow from the launch

Leaderboard numbers in the report are the authors’ measurements on the evals they named. They are not a guarantee about your documents, your language, or a later distilled model.

If a vendor says “V3-class,” ask which weights. A class is a marketing noun.

A hash is a file.

A reader who needs a decision should separate three things: the paper, the weight file they can download, and the API they pay for. Those can diverge.

Write which one you used.

What to record before you rely on it

Record the report URL, the model id or hash you actually ran, and the license file that shipped with it. If you quote a benchmark, name the eval and the fact that the authors measured it.

Keep the model id or the paper citation, the date you read it, and the inputs you used beside that note, so a later change can be compared with the same test.

When the result is a failure, store the artifact and change the design. Do not relax the check until the model happens to pass.

Name a person who owns that failure.

Which artifact you are talking about

When you write about DeepSeek-V3, say whether you mean the technical report, a weight file you can hash, or an API model id. Those three can move on different days.

A benchmark in the report stays a report number. A behavior you saw in a chat stays a behavior of that endpoint on that date.

Keeping the sentences in separate paragraphs is what stops a later reader from treating a family name as a single, measured thing.

Common questions

Does mixture-of-experts mean a cheaper API?

It can mean less compute per token than a dense model of the same parameter count. The price you pay is whatever the provider lists for that endpoint.

Check the price page. Do not infer it from the architecture.

Is every DeepSeek model V3?

No. The lab has published other models, including reasoning models and smaller files. Read the id.

The family name is not the checkpoint.

Sources & further reading

  1. DeepSeek-V3 Technical Report ↗
How this story was made

Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.

Our editorial standards
Back to all stories