Skip to content

v0.38.0 — Vision/Multimodal Support

Choose a tag to compare

@michaelneale michaelneale released this 20 Mar 07:12
· 1964 commits to main since this release

Vision/Multimodal Support

Models with vision capability now automatically get their multimodal projector (mmproj) downloaded and launched with --mmproj.

New

  • Vision support: Catalog models can declare an mmproj file. Download fetches it alongside the model. llama-server launches with --mmproj automatically.
  • 3 new vision models: Qwen3.5-0.8B, 4B, 9B (all vision-native)
  • Qwen3.5-27B tagged as vision-capable with mmproj
  • UI image attach: Button appears for vision-capable models. Attach images, see previews, send with messages.
  • 👁 badge: Vision models marked in model dropdown and dashboard
  • /v1/models capabilities: Returns ["text", "vision"] for vision models
  • /api/status vision field: vision: true/false per model

Changed

  • Mini swapped from Qwen3-8B to Qwen3.5-9B (same size, gains vision)
  • Fly API app (mesh-llm-api) retired — console serves both UI and API

Mesh