llama.cpp v0.5.0 (2026-09-23) accelerates CUDA conv2d via implicit GEMM, adds Metal MoE and SSM_CONV fusion, lets the server bind multi-address/--host lists and UNIX sockets, and adds HRM-Text (DFM Mimir 1B), MiMo-V2.6, and HunyuanOCR (DFlash) conversion plus ggml 0.25.0 backends.

Key Takeaways

  • ✓Shipped v0.5.0 on 2026-09-23 — major step past the prior v0.4.1 line
  • ✓Perf: CUDA conv2d implicit GEMM; Metal MoE + SSM_CONV fusion; CUDA graphs for MTP drafting
  • ✓Models: HRM-Text/Mimir 1B, MiMo-V2.6 convert, HunyuanOCR DFlash, Qwen4Exp sparse FA, Nemotron MTP/H
  • ✓Server: multi-address/--host lists + UNIX sockets; input_image on function-call outputs; router race fixes
  • ✓API: LoRA from FILE*; sampler env vars; llama-bench --version
🔬

In-Depth Technical Analysis

Core Background & Industry Pain Points

Local/edge stacks need CUDA and Apple Silicon plus new vision/MoE arches. Pain: slow conv2d/MoE kernels, single --host binding, lagging converters. llama.cpp v0.5.0 (2026-09-23) advances backends, model coverage, and server/router robustness with ggml 0.25.0.

Architecture Highlights & Internals

CUDA conv2d via implicit GEMM; Metal MoE + SSM_CONV fusion; CUDA graphs for MTP drafting. Server --host accepts comma-separated TCP addresses and UNIX sockets; function-call outputs may include input_image. New convert/arch paths: HRM-Text/Mimir 1B, MiMo-V2.6, HunyuanOCR DFlash, Qwen4Exp sparse FA, Nemotron MTP/H. API: LoRA from FILE*; sampler env vars.

Authoritative Benchmarks & Measured Scores

No official tok/s table in the notes — re-run llama-bench --version on your hardware. Multimodal correctness fixes include SigLIP bucket overruns.

Developer Hands-on Guide

Checkout v0.5.0, rebuild, and test multi-bind --host. See github.com/ggml-org/llama.cpp.