llama.cpp b9980 fixes multimodal detection for disabled mmproj
New release prevents false multimodal declarations when mmproj is explicitly disabled.
The llama.cpp team released version b9980, addressing a key multimodal handling issue. Previously, the server endpoint could incorrectly declare a model as supporting image/audio inputs even when the multimodal projection (mmproj) was explicitly disabled via preset or command-line parameters. The fix ensures that if `--no-mmproj-auto` is used, the model is not listed as multimodal on the `/v1/models` endpoint.
This release continues llama.cpp's focus on efficient local AI inference, offering builds across macOS (Apple Silicon, Intel, iOS XCFramework), Linux (CPU, Vulkan, ROCm, OpenVINO, SYCL), Windows (CPU, CUDA, Vulkan, OpenVINO, SYCL, HIP), and Android. The fix matters for developers running custom LLM stacks who need accurate API responses about model capabilities.
- Fixed incorrect multimodal declaration when mmproj is explicitly disabled via `--no-mmproj-auto`
- Affects `/v1/models` endpoint responses for models without image/audio support
- Supports builds across macOS, Linux, Windows, Android with multiple compute backends
Why It Matters
Ensures accurate API reporting for local AI models, preventing false multimodal capability declarations in production.