llama.cpp adds support for LiquidAI's d1-3B decision model
The b11483 release renamed internally to lfm2-d1 brings a specialised "decision" model type to the inference engine.
The facts
- llama.cpp release b11483 adds support for the LiquidAI/d1-3B decision model (PR #30110)
- The mtmd component now reads the LFM2 image resize algorithm from the GGUF file
- The decision type originally named 'd1' was renamed to 'lfm2-d1' in the common code
- Build artefacts span macOS, Linux (including CUDA 12/13, ROCm, Vulkan, SYCL, OpenVINO), Android, Windows and openEuler targets
llama.cpp's latest pre-release, tagged b11483, adds support for LiquidAI's d1-3B model. The project's commit log describes it as a decision model, a 3-billion-parameter system distinct from the chat models the engine usually ships.
The work landed via pull request #30110. It touches both the model-loading code and the multimodal (mtmd) component. That component now reads the LFM2 image resize algorithm directly from the model's GGUF file, rather than assuming a fixed setting.
A follow-up commit renamed the internal decision type from the terse 'd1' to 'lfm2-d1'. This ties the naming more explicitly to LiquidAI's LFM2 model family. Both commits in the release notes are tagged as assisted by Claude Opus 5.5, continuing llama.cpp's practice of crediting AI-assisted contributions in its changelog.
As with every tagged llama.cpp release, b11483 ships prebuilt binaries across a wide matrix. These include macOS on Apple Silicon and Intel, iOS, and several Linux CPU and GPU backends: CUDA 12.8 and 13.4, ROCm 10.0, Vulkan, SYCL and OpenVINO. Windows builds cover CPU, CUDA, OpenCL and Vulkan, alongside Android and openEuler on x86 and aarch64.
Why it matters
Decision-type models are a distinct architecture from the chat models llama.cpp usually ships. The GGUF-driven image resize config suggests LiquidAI's multimodal LFM2 line is getting first-class treatment in the runtime, not a bolt-on. Teams already self-hosting LFM2 variants should pull the new build before adopting d1-3B. The resize behaviour now depends on metadata baked into the GGUF file, not a runtime default.
Primary source
llama.cpp releases — github.comWritten from the primary source above by The Neural Net’s automated desk (claude-sonnet-5, theneuralnet/pipeline@1) without human review. How this works. Corrections: check the source first — if we got it wrong, the source link is the record.