Starstruk Logo MarkStarstruk

v0.6.2

tashfeenahmed/freellmapiv0.6.2Jul 28, 2026by tashfeenahmed

AI Summary

This release enhances proxy functionality with SOCKS support and environment variable detection, improves routing health by preventing synchronized bursts, and introduces editable model metadata. It also adds a new file upload feature to the playground.

Key Highlights

  • SOCKS proxies (socks5h/socks4a) and standard environment variables are now fully supported.
  • Health checks are optimized to interleave across providers instead of firing synchronized bursts.
  • Retired models (HTTP 410) are now auto-quarantined with a badge rather than consuming fallback slots.
  • Model metadata, including intelligence/speed ranks and RPM limits, is now editable in the dashboard.

New Features

  • Playground file upload for images and text files with client-side downscaling.
  • Custom endpoints 'Fetch models' to bulk-register newly discovered models.
  • Editable intelligence/speed ranks and rate limits per model in the dashboard.

Full Release Notes

A second release on top of v0.6.1, focused on proxy support, routing health, and making model metadata editable.

## Fixes

- **SOCKS proxies with remote DNS now work** — `socks5h://` and `socks4a://` are accepted, and the standard `ALL_PROXY` / `HTTPS_PROXY` / `HTTP_PROXY` / `NO_PROXY` environment variables are now detected (explicit `PROXY_URL` and dashboard settings still take precedence) (#630, #353, #643)
- **Cloudflare Workers AI streams no longer truncate mid-response** when the client omits `max_tokens` — a sensible default output budget is applied per provider (#553, #644)
- **Key health checks no longer fire synchronized bursts at one provider** — probes interleave across providers with per-provider spacing and interval jitter (#553, #644)
- **Chronic timeouts now lower a model's speed score** instead of only its reliability, and observed speed ranks feed the speed-sort presets (user-set overrides always win) (#619, #645)
- **Models retired by the upstream (HTTP 410 / end-of-life) are auto-quarantined** with a "retired" badge instead of burning a fallback slot on every request; a later catalog re-listing reinstates them (#634, #647)
- **Provider "not allowed to access model" errors (400/401 bodies) are now treated as access denials** rather than transient failures, ending an infinite cooldown-retry loop (#618, #647)

## New

- **Playground file upload** — attach images (with client-side downscaling) and text files; non-vision models are flagged while images are staged (#325, #648)
- **Custom endpoints: "Fetch models"** discovers the models your relay actually serves and bulk-registers them; newly added custom models are seeded at sensible ranks so auto-routing considers them (#488, #649)
- **Intelligence/speed ranks and RPM/RPD/TPM/TPD limits are now editable per model** in the dashboard, with per-field override indicators (#551, #646)

Full diff: https://github.com/tashfeenahmed/freellmapi/compare/v0.6.1...v0.6.2

Thanks to everyone who reported issues.

> **macOS note:** the DMG is not notarized — on first launch use right-click → Open (or `xattr -d com.apple.quarantine /Applications/FreeLLMAPI.app`).

---
⭐ Like the free router? [Go Premium](https://freellmapi.co/?utm_source=github&utm_medium=release&utm_campaign=premium&utm_content=v0.6.2#pricing) — the live signed catalog, $19/yr, cancel anytime.