SessionsAI

Cancel Actually Cancels - Stop a model while it loads: Cancel now interrupts loading a multi-gigabyte model, not just the reply that follows it. - Stop a long prompt: Cancelling during the "reading your prompt" stage takes effect immediately instead of waiting for it to finish. - Stop an image mid-draw: Image generation now aborts partway through and gives the memory straight back. Cancelling is treated as something you meant to do — no error alert, no Retry button. You Can See What It's Doing - Loading progress for big models: A large model now shows "Loading model… 42%" instead of "Thinking…" for minutes on end. More Room to Think - 1.5× more context on 8 GB iPhones and iPads: 8192 → 12288 tokens, from two memory corrections rather than optimism — enough for noticeably longer documents and conversations. Better Answers from the Same Models - Models now follow their own recipe: A model can carry the sampler settings its author recommends inside the model file, instead of every model sharing one profile. - Correct prompt formatting: The special tokens that structure a conversation are handled as tokens rather than as ordinary text, which is how the models were trained to see them. Image Generation, Corrected - Your photos keep their shape: Image-to-image preserves the subject's proportions instead of stretching everything to a square, and respects the orientation your phone recorded — portraits no longer come back rotated and cropped. - Reproducible images: The seed appears with the result, so you can recreate an image or vary it deliberately. - Smoother previews: Live previews no longer run a full decode on every single step. A Smaller, Newer Model Catalog - Gemma 4 E4B joins, alongside Gemma 4 E2B — and Gemma 4 is now Apache-2.0 licensed. - Older models retired: Gemma 3, Phi-4 mini, Ministral 3 8B and Mistral Nemo Prism are no longer offered to new installs. Anything you have already downloaded keeps working; retired models stay in your list until you delete them. - 12 text models offered, down from 16 — chosen so each one covers a device class the others do not. More Reliable Tools - Tool parameters from MCP servers now carry their descriptions and numeric limits through to the model, union-typed and combined-schema parameters are no longer dropped, and a tool that takes no arguments no longer has one invented for it. Privacy & Stability - Prompt-injection defences now derive their markers from the model's own vocabulary rather than a fixed list. - A build-time guard fails the release if the shipped image engine is not the exact version the build pins, and a data race in image-model downloads is fixed.