Backend Setback and Recovery

A short update, overdue.

Over the past few weeks we lost roughly four weeks of backend progress to a faulty git merge that was then imported into production. Code, configurations, working features — gone in a way that took days just to fully assess.

If you used AILinux during this period, you likely felt it. Models disappeared from pickers, image generation broke or returned garbled output, response quality regressed, and some endpoints returned errors they shouldn’t have. The disruption was real, and we know it was confusing for the people who quietly waited it out as much as for the ones who reached out.

We went quiet because we were buried in recovery, not because the project stalled. Rebuilding what was lost — without re-introducing the same bugs that had survived the merge — took longer than expected.

Where things stand now: the TriForce backend is healthy again. Image generation is fully restored across Replicate (including FLUX 1.1 Pro Ultra), OpenRouter (Google’s Nano Banana / Gemini image models, OpenAI GPT-5 Image), Cloudflare Workers AI, and the direct Gemini provider. The model registry is rebuilt. Federation across our three nodes is stable. We’re closing in on parity with the pre-incident state, commit by commit.

Still on the list for the coming days: HuggingFace integration with the right free-tier models, broader video generation support, and some tier-policy tuning so the right models reach the right users. We’re shipping fixes daily.

Thanks to everyone who stuck around — especially those who didn’t say a word but stayed anyway. We’re nearly back.

— Markus