Model comparison · 2026-10-19

EU-Hosted LLM Models: Reading a 57-Row Catalog

See what the 57 rows in the EU-hosted category offer, which capabilities were actually measured, and how to choose a row for the task you have in front of you.

What the category covers

The EU category holds 57 rows: 51 chat rows, 4 embedding models, 1 reranking model and 1 decision model. All of them are called with a single API key and no separate credential; the chat, embedding and reranking rows work with your existing OpenAI-compatible client, and the decision model uses the `/v1/systemone` endpoint. These counts include rows marked as retired in the catalog; a retired row cannot be called.

For the models in this category, processing inside the European Union is what the provider states; LLMTR has not independently verified it. The category comes from three providers, GreenPT, TensorX and Heabsy; each needs its own data-transfer consent, and you assess each provider separately.

Measured capabilities, not copied ones

The capability information in the catalog was not taken from the provider's marketing page — every GreenPT and TensorX row was run and tried before it was listed. In this category that distinction is not theoretical: one provider's own model card listed two models as image-capable, and in testing those models never processed an image.

Across those 49 chat rows the measurements came out as follows, and the catalog's capability lists were written from them. The capabilities of Heabsy's 2 rows are, for now, the provider's own statement and are not counted below.

Tool calling works on all 49 chat rows, verified over a full two-turn loop with the tool result consumed.
24 rows genuinely process images; the image capability appears on those rows only.
28 rows return cache hits on repeated prefixes, and all of them carry a discounted cache rate.
Response streaming works on all 49 chat rows, with complete usage in the final frame.

Choosing by task type

Rows in this category are not interchangeable. Among the measured rows, 37 of the chat rows reason whether you ask them to or not, and those hidden tokens bill as output; since output costs two to six times input on most rows, that decides the bill even on short requests.

For batch work where output cost has to be predictable, prefer the twelve rows that spend no reasoning tokens. If you need a row that reads images, pick one of the 24 that carry a measured image capability; send an image to a model without it and the request never reaches the provider — LLMTR refuses it without charging you.

Four embeddings and one reranker for retrieval

The category has four embedding rows, with measured output sizes of 2560, 4096 (two rows) and 3584 dimensions. Prices differ row by row; the current price is on each row's model page. Measure on your own collection which works best, because we ran no retrieval-quality comparison between these rows. Build a collection on one embedding row; changing rows means re-embedding the collection.

The reranking row scores and orders the candidates an embedding search returns. Evaluate ranking quality on your own set as well: in the one sample we have, on a three-document set, the model gave its highest score to a document unrelated to the query. That single observation is not a benchmark, but it is reason enough not to ship it without comparing against your own baseline.

Where to start

Turn on the European Union filter in the catalog; only this category's rows are listed and the filter appears in the address bar, so you can share the link with your team. A row's detail page shows its measured capabilities, context window and data policy note together.

Keep the first trial narrow: one task type, one row, and your own acceptance criterion. If the result meets it, widen the workload; if it does not, try a second row on the same task. Moving between rows in this category is a change of model identifier and nothing else.

The TensorX and Heabsy rows added to the category

On 2026-10-09 the category gained TensorX's 12 chat rows and 1 embedding row. Their ids have the form tensorx/<model>-tensorx, and they are called on the same endpoint with the same LLMTR key as every other row. They need the TensorX data-transfer consent; a consent granted for GreenPT does not cover them.

On 2026-10-13 Heabsy's 2 chat rows were added: Qwen3.8 27B and its row that is free until 2 November 2026. Their ids have the form heabsy/<model>-heabsy, and they need the Heabsy data-transfer consent.

Some model families sit in the category at both providers: GLM-5.3, GLM-5.3 Flash, DeepSeek V4.1 Flash, DeepSeek V4 Flash 0731, Kimi K2.7 Code and Qwen3.8 27B. If a row in one of these families returns 503, 429 or another server error, you can move to the other row of the same family by changing the model id; the switch works in both directions and happens in your client, LLMTR does not do it by itself.

Frequently asked questions

Do all 57 rows come from the same provider?

No. The category comes from three providers, GreenPT, TensorX and Heabsy, and every row runs on the same LLMTR API key. The data-transfer consent is per provider, though: you grant each one separately in Settings and assess each provider separately.

What happens if I send an image to a model without the image capability?

LLMTR refuses the request before it reaches the provider, and you are not charged. The error code is unsupported_input, so it says plainly that the model does not accept images.

Do I need a separate API key for these rows?

No. You call them with your existing LLMTR key by changing only the model identifier. Apart from the decision model, which uses `/v1/systemone`, there is no separate endpoint either.