Debian Inference Portal

Debian launched the Debian Inference Portal, a self-service service that lets Debian contributors log in with their Salsa account, create API keys, and track spending on shared LLM inference through an OpenAI-compatible proxy at inference.debian.net. The service is funded by a fixed monthly pool of credits sponsored by European cloud provider Scaleway, with weekly budgets set as a soft limit for fair use, and exposes two preview models: Zhipu AI's GLM-5.2 (scaleway/glm-5.2) and DeepSeek's V4 Flash (scaleway/deepseek-v4-flash-0731), with DeepSeek V4 Flash recommended for most uses. GLM-5.2 does not differentiate cached input tokens, so re-reading long context is charged at the full rate, which the portal notes can get expensive in agentic software development.

The Debian Inference Portal is a self-service way for Debian contributors to use shared LLM inference. You log in with your Salsa account, create API keys, and track your spend — all through your browser. The portal provisions you on a shared, OpenAI-compatible inference proxy, so any tool that speaks the OpenAI API can use your key. It is a shared, fair-use service funded by a fixed monthly pool of credits. Your weekly budget is a soft limit, in place for safety reasons : it paces usage so the shared pool lasts for everyone, and you can renew it from the dashboard once it is spent. Please still avoid long or very heavy workloads. Scaleway https://www.scaleway.com is a European cloud provider owned by the Iliad Group https://en.wikipedia.org/wiki/Iliad SA the parent of Free , delivering public cloud, bare metal, AI and managed services. Among those services is Generative APIs https://www.scaleway.com/en/generative-apis/ , a serverless, OpenAI-compatible API for serving LLMs — you pay per token and don't manage any infrastructure. The models below are served through it: requests are proxied to api.scaleway.ai , and you don't need your own Scaleway account — the service holds the provider credentials and forwards your calls. This service is possible thanks to Scaleway , which kindly sponsors the Debian project with a monthly allocation of credits to cover the inference costs for contributors. | Model ID use this in model | Model | |---|---| | scaleway/glm-5.2 | Zhipu AI's GLM-5.2, a general-purpose model | | scaleway/deepseek-v4-flash-0731 | DeepSeek's V4 Flash, a fast, lightweight model | Both models are in preview at Scaleway. They are billed by token, but note that GLM-5.2 currently does not differentiate cached input tokens, so re-reading long context is charged at the full rate and can get comparatively expensive. Re-reading long context is typical in the context of agentic software development agents re-read the accumulated conversation on every step , so this adds up quickly. DeepSeek V4 Flash is recommended for most uses. The list above is not exhaustive — run the models endpoint with your key to see exactly what is currently exposed: curl -s https://inference.debian.net/v1/models -H "Authorization: Bearer $DEBIAN INFERENCE KEY" bubblewrap bwrap \ --ro-bind /usr /usr \ --symlink usr/bin /bin \ --symlink usr/lib /lib \ --symlink usr/lib64 /lib64 \ --ro-bind /etc/ssl /etc/ssl \ --ro-bind /etc/resolv.conf /etc/resolv.conf \ --ro-bind $HOME/.gitconfig $HOME/.gitconfig \ --bind $HOME/.config/opencode $HOME/.config/opencode \ --bind $HOME/.cache/opencode $HOME/.cache/opencode \ --bind $HOME/.opencode $HOME/.opencode \ --bind $HOME/.local/share/opencode $HOME/.local/share/opencode \ --bind $HOME/.local/state/opencode $HOME/.local/state/opencode \ --proc /proc \ --dev /dev \ --tmpfs /tmp \ --bind $ pwd $ pwd \ --chdir $ pwd \ --unshare-all \ --share-net \ --die-with-parent \ /path/to/opencode "$@" This is not a full containment: the sandboxed process still has the terminal and can reach local sockets that don't go through the filesystem abstract Unix sockets, TCP sockets, and so on . nm.debian.org . Things to know about keys: The proxy speaks the OpenAI API https://en.wikipedia.org/wiki/OpenAI : the classic Chat Completions interface POST /v1/chat/completions with a messages payload and the newer Responses API POST /v1/responses . The examples below use Chat Completions, which is the most widely supported. Use your key as a Bearer token and set model to one of the exposed models e.g. scaleway/deepseek-v4-flash-0731 . export DEBIAN INFERENCE KEY="sk-your-key" List the models available to you curl -s https://inference.debian.net/v1/models \ -H "Authorization: Bearer $DEBIAN INFERENCE KEY" A simple chat completion curl -s https://inference.debian.net/v1/chat/completions \ -H "Authorization: Bearer $DEBIAN INFERENCE KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "scaleway/deepseek-v4-flash-0731", "messages": {"role": "user", "content": "Hello "} }' python import os import requests key = os.environ "DEBIAN INFERENCE KEY" resp = requests.post "https://inference.debian.net/v1/chat/completions", headers={"Authorization": f"Bearer {key}"}, json={ "model": "scaleway/deepseek-v4-flash-0731", "messages": {"role": "user", "content": "Hello "} , }, resp.raise for status print resp.json "choices" 0 "message" "content" Add a custom OpenAI-compatible provider to your opencode.json project-level, or the global ~/.config/opencode/opencode.json : { "model": "debian-inference/scaleway/deepseek-v4-flash-0731", "provider": { "debian-inference": { "npm": "@ai-sdk/openai-compatible", "name": "Debian Inference", "api": "https://inference.debian.net/v1", "env": "DEBIAN INFERENCE KEY" , "models": { "scaleway/deepseek-v4-flash-0731": { "name": "DeepSeek V4 Flash 0731", "family": "deepseek-flash", "release date": "2026-07-31", "attachment": false, "reasoning": true, "temperature": true, "tool call": true, "cost": { "input": 0.4, "output": 0.8, "cache read": 0.08 }, "limit": { "context": 256000, "output": 16384 }, "modalities": { "input": "text" , "output": "text" }, "interleaved": true }, "scaleway/glm-5.2": { "name": "GLM-5.2", "family": "glm", "release date": "2026-06-13", "attachment": false, "reasoning": true, "temperature": true, "tool call": true, "cost": { "input": 1.8, "output": 5.5 }, "limit": { "context": 256000, "output": 16384 }, "modalities": { "input": "text" , "output": "text" } } } } } } Then start OpenCode, use /connect , search for Debian Inference , and enter your API key. The example config above sets DeepSeek V4 Flash 0731 as the default model — use /models to switch to GLM-5.2 if you prefer it. Then make your first prompt. Pi https://pi.dev is a minimal, extensible agent harness. Add a custom OpenAI-compatible provider to your ~/.pi/agent/models.json to use the service: { "providers": { "debian-inference": { "baseUrl": "https://inference.debian.net/v1", "api": "openai-completions", "apiKey": "$DEBIAN INFERENCE KEY", "models": { "id": "scaleway/deepseek-v4-flash-0731", "name": "DeepSeek V4 Flash 0731", "reasoning": true, "contextWindow": 256000, "maxTokens": 16384, "cost": { "input": 0.4, "output": 0.8, "cacheRead": 0.08, "cacheWrite": 0.4 } }, { "id": "scaleway/glm-5.2", "name": "GLM-5.2", "reasoning": true, "contextWindow": 256000, "maxTokens": 16384, "cost": { "input": 1.8, "output": 5.5, "cacheRead": 1.8, "cacheWrite": 1.8 } } } } } The file reloads each time you open /model . Set DEBIAN INFERENCE KEY to your key for example via /login for the provider, or exporting the env var and pick a model with /model . Active Debian Developers and Debian Maintainers. Access is granted on sign-in, based on your status on nm.debian.org . Use it for work that benefits the Debian project — for example packaging, bug triage and fixes, tooling, documentation, or anything else you're doing as a contributor. It isn't for unrelated personal tasks. The underlying resources are provided free of charge by sponsors, so please be prepared to describe, in a short report, what you used it for if asked — such reports help make the case to sponsors for continuing and growing the service. Nothing directly — the service is provided for the Debian project and funded by a fixed monthly pool of credits from its sponsors. Spending is tracked per user over a weekly window that resets automatically, with defaults tiered by status: | Status | Default weekly budget | |---|---| | Debian Developer DD | $25 | | Debian Maintainer DM | $10 | This weekly budget is a soft limit, in place for safety reasons — not a hard cap on what you may use. It exists to pace usage so the shared monthly pool isn't drained by a few heavy users, which would leave nothing for everyone else. The limits are somewhat arbitrary and intended as a starting point; they may evolve over time as we learn how the service is used. Your current spend and the next automatic reset time are shown on the dashboard. Budgets are enforced by the proxy at the user level, so the total spend across all of your keys counts toward your limit. Once you reach it, new requests fail until you renew your credits below or the weekly window resets. Hitting the limit is normal for intensive work — it isn't a sign that you've done something wrong. Yes — renewing is the normal way to keep working once you've spent your weekly credits. The Renew credits button on the dashboard resets your weekly spend to 0, giving you a fresh window without changing your allowance. You can renew as soon as your credits are spent, subject to a safety check: because everyone shares one monthly pool, a renewal is only allowed while the predicted organization-wide monthly spend still fits comfortably inside the pool with a safety margin . The dashboard shows whether a renewal is currently available and the predicted monthly spend; when it is, use the Renew credits button. If it isn't available, wait for the weekly reset or contact the team https://salsa.debian.org/inference-team/inference-support/-/work items/new?type=Issue . scaleway/deepseek-v4-flash-0731 is the recommended default: it's fast, lightweight, and cheaper than GLM-5.2 because it differentiates cached tokens. scaleway/glm-5.2 is a strong general-purpose model, but it's in preview and currently charges cached input at the full rate. Use /v1/models to see what's currently available. Keys are shown only once at creation. Revoke it on the dashboard and create a new one. Check the expiry date on the dashboard and renew if needed. If you've exceeded your budget, renew your credits from the dashboard, or wait for the weekly reset. Yes — any tool that lets you set a custom base URL https://inference.debian.net/v1 and a Bearer token can use your key see the examples above . Outputs generated by the service belong to you. Scaleway does not claim any copyright or ownership over generated content. Scaleway applies a Zero Data Retention Policy : prompt content is not stored, and is explicitly not used for training, retraining, or improving the base models. Data is not accessible to the LLM creators or to third parties. Only anonymized metadata token counts, HTTP status codes is retained for up to 6 months for performance monitoring. Data is hosted in Paris, France, and as a European company Scaleway is not subject to extraterritorial laws such as the US Cloud Act. See Scaleway's Generative APIs privacy policy https://www.scaleway.com/en/docs/generative-apis/reference-content/data-privacy/ and Specific Conditions for AI Services https://www.scaleway.com/en/terms/ for the full details. Questions, feedback, or problems? Contact the Debian AI team by filing an issue at inference-support https://salsa.debian.org/inference-team/inference-support/-/work items/new?type=Issue .