FreeLLMAPI v0.11.1 is a fixes release: the desktop app stops offering updates that have no installer, the failover chain wastes less of its budget on dead routes and slow but healthy endpoints, and ACLIDE joins as a provider.
Desktop
- Update check. The desktop app now compares itself against the latest published release instead of
main. Before this, every untagged commit showed up as "Update available" with nothing to install behind it. (#1279, closes #1270) - Sign-in. The login form submits the desktop account identifier, so signing in from the desktop app works again. (#1266)
- Linux .deb.
chrome-sandboxships with mode 4755, so the app starts on distributions that enforce the sandbox helper's permissions. (#1240, thanks @rome-xi) - Releases are now published only once every installer is attached, so the release page never shows source archives alone while the builds run. (#1247)
Routing and failover
- Slow but healthy endpoints. The retry budget is now derived per endpoint from its own time-to-first-byte history,
max(base budget, weighted P95 + buffer), so a later failover hop to an endpoint that normally takes 50 seconds is no longer cancelled at the fixed 45. Fast endpoints and endpoints with no history keep the base budget. Override with settings orTTFB_BUDGET_DISABLED. (#1272, closes #1262, thanks @Inference1) - Models that reject tool calls are remembered per model and endpoint and tried last for requests that carry tools, instead of being walked first on every tool request. (#1267, closes #1230)
- Stale catalogs. After three distinct
model_not_foundmisses on one platform, the rest of that platform is skipped for the request and the next provider is tried. (#1222, thanks @suantea) - Truncated streams. A route that cuts its stream short three times in a row is benched for five minutes rather than the 90 second default. (#1221, closes #1218, thanks @suantea)
- Out-of-credits replies sent as 200. Upstreams such as Pollinations that answer 200 with a top-up notice as the reply now fail over like a 402, instead of the notice being returned as the model's answer. (#1241, thanks @sukaimi)
Providers
- ACLIDE. A dedicated Responses adapter with text and image input, tools, structured output, reasoning summaries and usage. Streaming is buffered. Keys import through the usual setup and share the monthly Free allowance. (#1273)
- Sail. Every model now uses the flex completion window. Sail rejects
background: truewithcompletion_window: asap, which made most Sail models fail with a 400. (#1259, closes #1248, thanks @Icesenator) - OpenCode and Reka. OpenCode Zen's free roster only answers from inside OpenCode, and Reka now needs prepaid credit, so neither is described as free any more. The OpenCode client headers from #1204 made no difference and were removed. (#1265, closes #1202 and #1249)
Updating
Download the installer for your platform below, or use ghcr.io/tashfeenahmed/freellmapi:v0.11.1 for Docker. The latest Docker tag also points to this release.
Desktop users on v0.11.0 need to install this one by hand from this page. From v0.11.1 on, the Updates panel only reports real releases.
Existing keys and profiles are retained. No new database migrations in this release. Back up your data before updating a self-hosted instance.
⭐ Like the free router? Go Premium, the live signed catalog, $19/yr, cancel anytime.