FreeLLMAPI: all versions and old releases

Zro by MoonMath.ai
- Add your own Zro API key in the Keys dashboard; imports recognize ZRO_API_KEY and MOONMATH_API_KEY.
- OpenAI-compatible normal and streaming chat, tool calls and JSON output use the existing routing and failover behavior.
- Credential validation uses the authenticated model-list endpoint without spending inference tokens.
- Reject mismatched response model identities, credential redirects and in-band streaming errors.
- Account-wide Free usage resets daily; it is not a monthly cash grant. Model availability is managed through the signed catalog rather than bundled seeds.
Catalog compatibility
This release supplies the Zro adapter needed by the November 12 catalog update. The existing Premium-immediate / Free-after-30-days catalog logic is unchanged. Mistral and other providers' application behavior are unchanged.
The catalog publication is separate from the app release. The validated update holds three dated/non-renewing promotions rather than mislabeling them as recurring free, retains stored history, and adds only tested routes.
Local validation: full server/CLI/client suites, locale validation, migration round trips, lint, workspace build and desktop tests. Zro was also tested through the real app router against the Free account with an in-memory database.

Outcrawl web search
- Add Outcrawl (api.outcrawl.ai) as a keyed web search engine (#1802). Its free plan gives 1,000 credits a month with no card. A search costs about 0.5 credit per result, so roughly 400 searches at the default 5 results. Key checks read the credit balance and spend nothing. News searches run as ordinary web searches. Thanks @heydryft.
- The engine arrives through the signed catalog: Premium gets it now, and Free gets it after the usual 30 days.
Routing
- Each chain can now have its own routing strategy (Manual order, Balanced, Smartest, Fastest, Most reliable) or inherit the global one (#1697, #1696). An auto:<axis> alias in the request still wins over the chain's setting. Thanks @navzo.
Fixes
- The monthly token budget bar keeps counting tokens after analytics retention deletes old request rows (#1782). History deleted before this update cannot be recovered. Thanks @kvianAR.
Maintenance
- Upgrade Zod to v4 and Vitest to v5 (#1779). Validation errors still return 400, now with Zod 4's wording. Thanks @thedevil11255.
- The CLI test suite passes on Windows (#1803). Thanks @leucds2018-hub.
Also in v0.16.25 (not listed in its notes)
- A model that answers "this model is overloaded / experiencing high demand" with a 5xx is skipped for the rest of the request instead of being retried on every key (#1781). Thanks @kvianAR.
- Gemini no longer rejects tool schemas that put items on a non-array schema (#1791, fixes #1790). Thanks @TanmayShah29.
- Settings → General has a "Start on system startup" switch on macOS and Windows (#1799). Thanks @MoriDanWork.

FreeTTS: local personal BYOK speech
- Add FreeTTS key setup and import support for your own local account.
- Add native MP3 text-to-speech with voice selection, fixed-host credential-safe downloads, bounded responses, audio validation and rate-limit failover.
- Keep the provider's spoken watermark. FreeTTS Free is personal/non-commercial; local use does not waive its terms or permit reselling shared API access.
- The signed catalog adds one logical TTS entry, neural-tts. Voices are parameters, not thousands of separate models. Premium receives the catalog entry now; Free receives it after the existing 30-day aging rule. No gate changes or early model seeding.
- Keep every other provider/model row, Mistral settings, and existing user overrides unchanged.
Install this release to use the FreeTTS adapter. Add your own FreeTTS key locally; no provider key is bundled or published.
The release also includes the fixes already merged on main since v0.16.24, including Gemini schema compatibility, overloaded-model sibling-key fallback, and the Start on system startup toggle.
Linux installer builds are pinned to Ubuntu 24.04 for compatibility with the current packaging tools; AppImage, tar.xz, DEB and RPM remain available.
Provider details: https://freetts.org/developers · https://freetts.org/pricing · https://freetts.org/terms

PowerTokens provider
- Add PowerTokens (api.powertokens.ai) as an OpenAI-compatible provider. Its glm-4.7-flash model is listed at $0 (0 Credits); chat, streaming and tool calling were tested through the app router.
- PowerTokens is often overloaded and answers 429 (code 1305). The router cools the key down and moves on, so treat it as an extra fallback rather than a primary route.
- Keys restricted to other models in the PowerTokens dashboard return 403 token_model_forbidden; allow glm-4.7-flash on the key.
The model arrives through the signed catalog. Premium receives the live catalog; Free receives the new row after the unchanged 30-day aging period. Thanks @PowerTokens for the original contribution (#1659).
Fixes
- Custom endpoints with several keys now rotate image, speech, transcription and embedding requests across those keys when one is disabled, out of budget, rate limited or rejected (#1738, fixes #1737). Thanks @w21180239.
- Free.ai is no longer offered for chat routing, and "Test" on a legacy non-chat model row explains that the platform does not support chat instead of reporting a missing key (#1769). Thanks @w21180239.
- "Probe now" gives reasoning models room for their hidden thinking (#1862).
- freellmapi setup-opencode stores the key in auth.json instead of an environment placeholder (#1861).
Docs
- Turkish README translation (#1805). Thanks @alkank.