Changelog

RAM calculator and AI coach only suggest measures the model type supports – for YOLO that is batch, imgsz and FP16. Apply buttons for settings without effect are filtered out.

Gradient checkpointing and group by length are hidden for YOLO.

The RAM estimate for image models uses the image size (imgsz) instead of the text length (1.2.61).

The Linux build no longer fails because of a single unreachable package server; Linux packages are back from 1.2.61.

Skills in the AI coach: typing / in the input field opens prepared questions for the current page, such as /ram, /assess or /review.

The RAM estimate is compared with your actual memory and reports “fits”, “tight” or “not enough” – in the panel and in the coach with the same formula (1.2.57).

Coach buttons only appear where they have an effect on the page; token usage per round drops by 52–65% (1.2.57).

A code suggestion that does not fit is marked “not applicable” instead of reporting success; applied code can no longer break the script (1.2.55, 1.2.56).

One API key per AI provider: Claude, OpenAI and Groq each keep their key, switching no longer loses one.

Every AI feature works reliably at every token budget; Claude's reasoning effort scales with the budget (1.2.48).

Empty AI answers show an explanatory error; the AI proxy has timeouts instead of loading forever (1.2.48).

Saving with an empty key field no longer deletes a stored key.