EuroLLM and Apertus on Workers AI: access requests, model IDs and context windows

An announcement, a model page and account access are different evidence

The official Cloudflare announcement on October 1 says two European models are coming to Workers AI and access can be requested that day. This article was checked on October 2; that is the review date, not a second model-release date.

Both model pages list Cloudflare-hosted deployment, parameters, context windows and Usage examples, while explicitly requiring an access request. They identify the model names, application route and preparation steps, but do not establish that a reader's account is approved. They also do not provide a complete GA declaration, approval schedule or region-by-region availability table.

Copy the exact model ID before comparing context windows

The current EuroLLM model page lists <code>@cf/utter-project/eurollm-9b-it</code>, 9B parameters and a 32,000-token context. The current Apertus model page lists <code>@cf/swiss-ai/apertus-v1.5-8b</code>, 8B parameters and a 262,144-token context. Do not substitute a blog's family name, another version or another platform's ID in a Workers AI request.

A context window is not the maximum output length, nor a guarantee of accurate answers for all long inputs. Test input and output limits, error messages, latency and cost. This article does not turn the two window sizes into an unmeasured performance comparison. The pages also label function calling, and Apertus labels vision, but those labels do not prove that your account has the corresponding access or that existing tool-call and image formats are compatible.

Apply through the official model page and retain the approved scope

Both model pages link to the same access request form. Follow the link from the official page. An authorized owner should review the requirements and what information the organization permits submitting, and describe the models and scenarios to evaluate. Do not submit production data or credentials without checking the purpose. This site did not open and submit the form, fill it in for anyone or connect any account.

Keep the application record and official reply, confirming the approved account, models, capabilities and actual limits. No standard approval time was verified, so immediate access cannot be promised. A successful request submission or public documentation is not approval. Without explicit access confirmation, wait or check through official channels rather than repeatedly changing model IDs, creating accounts or switching IPs to infer eligibility.

After approval, validate a minimal text request from the documentation

Both pages show the Workers AI binding form <code>env.AI.run(modelId, { messages })</code> and a REST path. In an existing Workers AI test environment, confirm the AI binding, account and authorization, then use the page's exact modelId with one nonsensitive text input. Do not overwrite a production Worker with the complete example.

Start with a non-streaming text request and record the ID, test time, input language, return status and response. Test streaming, long context or tool calls only after the basic path succeeds. Retain error types and check access permissions, model spelling, binding/authentication and request format separately. Playground and OpenAI-compatible endpoint references are documentation entry points; they do not independently prove account access to a new model or that every endpoint has been tested.

This article provides a procedure based on official examples. It did not run inference, measure throughput or demonstrate success for a particular account. For the distinction between keys, base URLs and model names, see the API configuration guide. Network configuration cannot replace Workers AI access approval.

Separate multilingual coverage claims from acceptance results

The official announcement says EuroLLM supports 35 languages, including all 24 official EU languages, and that Apertus training covers more than 1,500 languages. These are vendor descriptions of coverage or training, not per-language accuracy tests performed by this site. Training coverage does not mean equal capability in every language.

Choose the languages and tasks you actually need, then prepare a small set of samples people can assess: ordinary questions, terminology, regional expressions, mixed languages and unknown questions the model should decline. For long documents, check whether key information can still be located accurately at the beginning, middle and end, with review by someone who reads that language. Compare models against the same acceptance criteria, not parameter counts or window sizes alone.

European model provenance does not establish data residency

The announcement discusses AI sovereignty and model choice, including development by public universities and research institutions. The Workers AI pages state Cloudflare-hosted. Model provenance, open training data and the hosting platform are different facts. These three materials do not directly promise that requests for these models are processed only in the EU or Switzerland, follow a particular retention policy or meet a specific compliance requirement.

If a use case requires particular processing regions or retention rules, confirm applicable formal documentation and contractual terms before integration, then decide permitted inputs through the organization's data-approval process. Do not treat sovereign, European, open or a model name as a location guarantee, or public training data as evidence that your requests will not be retained.

Decide the next step with pricing and quotas still unconfirmed

The model pages reviewed do not specify separate pricing, account rate quotas, maximum output or approval times for these access-request models. Pricing and Limits links in navigation do not prove that a particular standard rate applies. After approval, check the applicable formal terms and actual account limits rather than inventing free or unlimited-use promises.

Teams with a multilingual need that can start with a limited evaluation may prepare acceptance samples and request access. If immediate production use, definite costs or regions, or verified image and tool-call paths are required, first obtain that evidence. Record the gaps and follow model pages and official notices; a new model announcement is not a reason to skip access and quality acceptance.