Ollaya 0.7.5 adds decider:2b-vision and image inputs
On 28 September 2026, Ollaya shipped v0.7.5: pull decider:2b-vision (Mapika’s vision-language Decider-2B on Qwen3.5-2B-VL) and pass images through the CLI (--image) or HTTP (images base64), alongside the usual typed questions over a text state.
The release notes claim pixel/token parity with the model’s own preprocessing on a 404-question check and about 200 ms for five questions on a 256×240 frame on an RTX 4090 (~0.9 s at 1000×750) — all UnverifiedClaim. Limits called out by the maintainers: one PNG per request, roughly up to 1 MP, and the usual “not affiliated with Ollama or TypeSafe” disclaimer.
Ollaya remains a runtime (no trained weights of its own). Stars moved roughly 547 → 786 between the 27 Sep and 28 Sep radar checks. Runtime entry: /runtimes/ollaya. Related model weights: Mapika/decider-2b-vision.