Skip to content
← Insights
Nov 14, 20264 min read

The model family name is missing half the specification

Compare exact releases, access routes and licences. A hosted app, an API and downloadable weights can be different systems under the same family name.

A pastel model archive box beside a cloud symbol, a licence document and a small desktop computer.

A model family name is too broad to put into a production specification. The same supplier can publish downloadable weights, operate a hosted API, and run a consumer app with different capabilities. A good result in one of those settings cannot settle the others.

This is especially easy to miss in lists organised by country. Qwen, DeepSeek, Hunyuan, and other Chinese model families deserve evaluation on the actual task. Their place of origin does not make their releases interchangeable, or turn a vendor comparison into an independent test.

A useful shortlist should name the exact release, the way it will be accessed, and the terms that apply. That record gives the next person enough information to reproduce the comparison.

Record what is actually available

Qwen's original Qwen3 announcement described particular open-weight models under Apache 2.0. That is a statement about those releases. It does not assign the same licence to every later product carrying the Qwen name.

The August 2026 Qwen3.8-Max announcement illustrates another distinction: a release announcement can include a plan to publish weights later. A team planning a local deployment needs the actual checkpoint and its licence before treating that plan as a downloadable dependency.

For each candidate, keep the official model card or release page with the selected revision. Record whether the test used a quantised checkpoint, a hosted endpoint, or a consumer interface. Quantisation and serving settings can change the system being measured even when the family name stays the same.

The record should include the date. A reproducible comparison needs a way to identify what was tested after a supplier changes its default endpoint or replaces a product page.

Separate the app from the API

DeepSeek's current change log discusses its API releases separately from the app experience, where search and other capabilities work together with inference. On November 13, 2026, it announced application APIs for capabilities that include search and file input. The supported model and parameters are specified in that release.

An answer produced with search is evidence about that combined system. It cannot be attributed solely to a downloadable model with no retrieval tool. Likewise, a consumer application's file upload does not establish that the same operation is available under an API contract.

Write the needed operation before choosing a provider. If the task is extracting fields from supplied documents, the comparison needs the permitted document input and the required output. If it involves calling tools, test the actual tool interface and failure behaviour.

This keeps a visually impressive demo from deciding an unrelated integration. The production task may depend on a capability the demo supplied through an additional service.

Read the licence for the selected artifact

Open weights describe access to model parameters. They do not by themselves specify commercial permissions, redistribution requirements, or geographic conditions. The selected licence is the source for those questions.

For example, HunyuanVideo's published licence contains its own conditions, including a separate permission requirement above a stated user threshold. It should be read for the artifact being used. A permissive licence attached to one text model cannot be transferred to a video model from another supplier.

Code dependencies can have separate terms too. Record the inference implementation and any additional components required to run it. A model download is only part of the software a team may distribute or operate.

For a client project, someone needs to own that review. A benchmark position says nothing about whether the planned deployment fits the licence or the client's data-handling requirements.

Compare one real workload at a time

For text, retain representative prompts and reviewer expectations. For image or video work, specify the intended output and the corrections an editor would need to make. For speech, use authorised voices and test the languages actually required.

Count failed attempts and review effort alongside successful samples. A low price for one generation can become a higher delivery cost if several attempts are discarded. A hosted service also introduces a different operating burden from a local checkpoint.

The outcome should be a dated decision about one task: which exact system passed, which conditions remain unresolved, and what evidence justified the choice. That is a model map someone can use. A list of family names with permanent rankings will start losing its meaning as soon as the next release changes what those names contain.

AIOpen modelsEngineering