v0.3.0 (unreleased)

FilesCopy Anchor Link

• An encrypted file store. The gateway can now keep its own files, on local disk or S3-compatible storage (Amazon S3, MinIO, RustFS, Cloudflare R2 and others). Every object is encrypted with its own key under rotatable master keys before any backend sees it. Off by default. See File storage.
• The Files API. /v1/files is OpenAI-compatible: upload, list, retrieve, download and delete, with workspace-scoped gateway ids and uploads streamed straight to the store. See Files.
• Workspace › Files: a page to upload, download and delete files.
• Storage quota. A new Storage limit per workspace (1 GiB by default, a platform override, tighter local limits), checked while uploads stream. Storage is recorded in GB-days and shown as Not charged.
• Admin › Settings › Data & privacy › Storage: the backend, encryption key and health, a storage test, and which kinds of file are allowed and for how long.
• New commands: files verify and files sweep --once.

Batches for any modelCopy Anchor Link

• /v1/batches now reads gateway files and runs Chat Completions, Responses, Embeddings and Messages lines, for any model.
• The whole input is checked first, with a report by line number.
• Native on OpenAI Batch and Anthropic Message Batches when every line uses one route; gateway-run otherwise, line by line with exactly-once claims and resume after a restart.
• Batch prices: routes can publish the provider's batch rates beside their standard prices. They are never worked out automatically.
• Results and errors are encrypted batch_output files. Logs › Batches and a batch page show progress and cost; new Batch failed and Batch stalled alerts and batch metrics.

See Batches.

Batch scheduling for self-hosted modelsCopy Anchor Link

• Gateway-run batch lines start only when their route has capacity: a per-route limit on lines at once (2 by default), and optionally yielding to live traffic, the server's load from vLLM-compatible /metrics on an approved endpoint (failing closed), and time windows. Routes are shared fairly between workspaces and batches, and a line stays on the route it was scheduled on.
• An optional vLLM priority hint on batch lines (vLLM and OpenAI-compatible routes) keeps them behind live requests inside vLLM.
• Batches show Queued — waiting for capacity with the reason, and legitimate waiting doesn't count toward Batch stalled alerts.
• completion_window can be 24h, 48h, 72h or 168h; anything but 24h always runs through the gateway. An expired batch keeps its partial results, and lines that never ran cost nothing.

See Batch scheduling.

LimitsCopy Anchor Link

• Jobs at once. A new limit on every layer: how many batch and video jobs may be active at once. Type defaults allow 2 per workspace. Jobs no longer hold "requests at once" slots and skip per-minute limits, but stay fully budgeted. A refused job gets 429 job_limit_exceeded.

SCIMCopy Anchor Link

• The last Platform Admin is protected. SCIM can no longer deactivate or remove the last active Platform Admin: the whole request is refused with 409, recorded in the audit log, and raises a built-in alert. See SCIM.

RealtimeCopy Anchor Link

• Holds sized from the conversation. Each response's hold is now worked out from the session's reported context plus new input, up to the model's context window, instead of the full input ceiling every time. Responses cut off at their output limit show as Incomplete.

VideoCopy Anchor Link

• OpenAI shut down its Videos API on 2026-09-24. /v1/videos now returns unsupported_capability, and the dashboard says so. An OpenRouter video adapter is planned.

Releases and securityCopy Anchor Link

• Versioned, signed images. The first release published as a container image: ghcr.io/ncecere/open-model-gateway:v0.3.0, also tagged v0.3 and latest-release, signed with cosign keyless signing. See Container and Compose.
• The binary's version is now 0.3.0 (it reported 0.1.0 before), as shown by gateway_build_info. The release notes are also in the repository, in docs/releases/v0.3.0.md.
• A security policy. SECURITY.md in the repository says how to report a vulnerability privately.
• The project is now under the MIT licence.

UpgradingCopy Anchor Link

Migrations 0018_job_limits, 0019_file_store, 0020_files_api, 0021_batch_engine and 0022_batch_scheduling. Back up, run migrate, reapply runtime grants, then review the new "jobs at once" defaults and, to use files and batches, configure the file store. See Upgrades.