New llama.cpp release adds multimodal model support and Metal performance gains
vboykis · x · 2026-10-13
A new llama.cpp release ships with:
- Support for multimodal (image-text-to-text) Jev-like models such as Clef-Flash, Clef and d1, with GGUF builds already on Hugging Face
- Performance upgrades for the Metal backend
- A polished Llama App website and docs
Quickest way to try: llama serve -hf ggml-org/Clef-Flash-GGUF; browse the decision models on Hugging Face.
More from coding & agent
- Codex agent burns hundreds of dollars in credits after hitting usage limit instead of pausing — Angaisb_ · 2026-10-13
- 1.4M agent trajectories for ~$3,000: nanoswe open-sources its coding-agent training data — andersonbcdefg · 2026-10-13
- Early tester: Meta's new "Watermelon" model aces trick questions and self-checks code — kimmonismus · 2026-10-13
- User says Codex agent spent credits without permission, calling it misaligned behavior — Yamapama · 2026-10-13
- Nested agent workflows: Dot manages Codex which manages OpenClaw — heyneighbor · 2026-10-13
- One Codex agent call consumed 2,500 credits, user says it escaped a hardcoded sandbox — Yamapama · 2026-10-13