Debian has launched a "Debian Inference Portal" that lets contributors use large language models. Debian Developers (DD) and Debian Maintainers (DM) can log in with their Salsa accounts, create an API key, and access models provided by Scaleway. Debian only recently approved conditional use of generative AI in contributions through an August vote. What has now been set up is not an AI service open to everyone, but an entry point for contributors to use in development while managing shared inference costs.
Connecting Debian's authentication to Scaleway's inference
The portal verifies users and manages API keys and budgets on Debian's side. Actual requests to the models are forwarded to Scaleway's "Generative APIs." At login, the system checks Debian's developer credentials database to confirm the user holds valid DD or DM status. Simply having a Salsa account does not qualify someone for access. Users do not need to create a Scaleway account themselves. Debian's service holds the provider credentials and relays calls made by contributors.
Two models are currently available: DeepSeek V4 Flash and GLM-5.2, both listed as preview status on Scaleway. The portal's documentation recommends DeepSeek V4 Flash for general use, citing faster processing and lower cost. However, this is guidance from the service provider's side, not an independent quality comparison based on Debian's own work. Model availability may change, so users need to check the model list after authenticating.
The API is OpenAI-compatible and supports both the conversational "Chat Completions" and "Responses" formats. Any development tool that lets you specify an endpoint can call the API using the key. The official documentation includes configuration examples for OpenCode and Pi. That said, Debian is not distributing these tools itself. The documentation explicitly notes that modern coding agents are not packaged for Debian.
Free, but not unlimited
To cover inference costs, Scaleway provides the Debian Project with a fixed monthly credit allowance. Contributors don't pay directly, but everyone draws from the same pool. DDs and DMs have different default weekly usage quotas, and both are subject to the shared monthly budget and renewal conditions. The initial settings are as follows:
| Eligibility | Default Weekly Quota | Applies To |
|---|---|---|
| Debian Developer (DD) | US$25 | Contributors with valid DD status |
| Debian Maintainer (DM) | US$10 | Contributors with valid DM status |
This amount is not a cash grant; it's a quota for tracking each user's inference request costs. If the same person creates multiple API keys, their usage is combined. Once the quota is reached, new requests will fail.
The weekly quota resets automatically. After it's used up, users can also reset their weekly spending to zero using the "Renew credits" button on the dashboard. However, pressing the button doesn't guarantee an immediate reset. It only works if the projected monthly spending across all users still fits within the shared pool with a safety margin. The weekly quota is called a "soft limit" precisely because using it up doesn't necessarily mean you're completely done. On the other hand, if the shared monthly funds run short, the renewal won't go through, and the user will need to wait for the next weekly reset or contact the maintainers. The initial amounts themselves may also change based on usage patterns.
Key expiration and the cost of long conversations
API keys are shown only once, at creation time; if lost, the existing key must be revoked and a new one created. New keys default to a 90-day validity period. Renewing through the dashboard extends this by another 90 days. This is a separate operation from "Renew credits," which resets the cost quota. Keys no longer in use can be revoked.
Model choice also affects the shared budget. According to the official documentation, GLM-5.2 does not offer a separate discounted rate for cached input tokens, so repeatedly reading long conversation histories incurs the standard input rate each time. Development agents tend to re-read accumulated conversation at each step. So even within the same usage quota, how much work can get done depends on which model is chosen and how long a context is sent, and how often. This is also why the portal recommends DeepSeek V4 Flash as the default option.
Debian's documentation also encourages sandboxing when letting an agent read a working directory and execute commands. It includes an example configuration using bubblewrap, which is available on Debian, while cautioning that terminals and certain local communication paths still remain accessible, so this doesn't amount to full containment. Even with the inference API entry point in place, users still need to manage the permissions of any agent running in their local environment.
Users send requests to the portal's API, but those requests are then forwarded to Scaleway's inference service. One should not assume everything is processed entirely within Debian and casually paste in confidential material. In personal development environments, users also need to check which files an agent is being given access to read.
Inference service use and Debian contribution rules are separate matters
Debian's General Resolution "Responsible Use of Generative AI" was adopted in August 2026. The project neither recommends nor prohibits the use of generative AI. Work produced with AI assistance is held to the same standards of quality and accuracy as before. Maintainability and legal compliance are likewise not exempted. Contributors are expected to understand the content they submit, review it, perform any necessary testing, and take responsibility for the results. Disclosing AI use is encouraged but not mandatory.
The resolution requires that private discussions, pre-disclosure security information, and credentials not be sent to third-party AI services without authorization. Automated processes with broad impact, such as mass bug reports or patch submissions, require prior discussion and human oversight. The availability of the portal should not be read as a relaxation of these conditions.
Use of the portal is also limited to Debian-related contributions. It can be used for packaging, bug triage and fixes, development tooling, and documentation work, but unrelated personal use is out of scope. To help explain continued support to sponsors, users may be asked to briefly report their use case. Regarding Scaleway's data handling, the official documentation states that prompts are not stored and are not used to train models, and that anonymized usage metadata is retained for up to six months. This is an explanation from the operator's side, and it does not provide grounds for expanding what information may be input beyond what Debian's resolution permits.
How much this new entry point will actually help contributors' work cannot be judged from undisclosed user counts or model performance alone. What each person should check first is which models are currently available, the remaining quota shown on the dashboard, and whether the data being sent is something that can be made public. If shared costs and contributor responsibility can be balanced, the inference service can become a practical tool supporting volunteer development work.