
The llama.cpp b12458 pre-release introduces support for running YuE2-3B inside the project's mtmd multimodal module. Tagged on GitHub as commit 92192d3 at 05:39 UTC on November 17, 2026 (late November 16 Pacific Time), the update bundles model compatibility with server concurrency fixes and backend graph optimizations.
What Changed in b12458
The centerpiece of the build is pull request #32369, co-authored by contributor Pascal, which brings YuE2-3B support alongside several runtime adjustments. According to the release notes, the update adds classifier-free guidance (CFG) with API-level CFG controls, exposes a flow_steps configuration parameter, fixes streaming and batch API behavior, enforces context-size checks, and resolves issues around n_predict.
Concurrency and Backend Graph Fixes
Two lower-level changes target correctness and execution efficiency in production and local server deployments:
- Per-slot logit isolation: Previously, context logits reflected the most recent decode across any active server slot. In multi-slot serving, YuE2 could sample its score and unguided codes from a different sequence. The updated
save_logits()path stores each sequence's own decode logits and performs the guidance mix there. - Snake activation chain fusion: A beta reshape operation previously sat between
sqrandmul, breaking the consecutivemul, sin, sqr, mul, addpattern and preventing compute backends from fusing the chain. Following the pattern used inqwen3tts, both reshapes are now expanded upfront so the math chain stays contiguous.
Availability and Practical Caveats
Pre-built binaries for b12458 span macOS Apple Silicon, iOS, Windows, Android, and Linux across CPU, CUDA 12 and 13, Vulkan, ROCm 10.0, OpenVINO, SYCL, and Snapdragon NPU/Adreno targets, while macOS Intel, KleidiAI-enabled macOS builds, and openEuler packages are marked disabled in the release matrix. Because b12458 is distributed as a pre-release and the changelog provides no comparative throughput or memory benchmarks, developers evaluating YuE2-3B will need to verify performance on their own hardware and source model weights separately under the upstream model's licensing terms.
EXPLORE THE SOURCE
Projects in this story
GitHub repositories
- ggml-org/llama.cpp ↗View project details
Automated Vertex AI (gemini-4-hydrogen) primary-source report; reviewed 2026-11-17.



