llama.cpp v0.7.0
Overview llama.cpp v0.7.0 adds the LiliCorr speculative decoding technique and MTP for GLM5-Next, a new GPU cache for MoE experts kept in host memory ( --moe-cache-mib , up to ~2.2
Virticle Desk · Edited to Virticle Standards
October 15, 2026
4 minute read
The human is the plot.
In short: llama.cpp v0.7.0
What moved
Overview llama.cpp v0.7.0 adds the LiliCorr speculative decoding technique and MTP for GLM5-Next, a new GPU cache for MoE experts kept in host memory (`--moe-cache-mib`, up to ~2.2x decode speedup when experts are offloaded), and a large batch of new models: K2 Horizon (dense and MoVA), Prism Bonsai 2 27B, Embedding
This cleared Virticle’s weekday bar because it looks consequential for someone who is not giving a keynote — not because it won a thread.
Why a human should care
Ask what default, power relation, or daily ritual actually changed. If the answer is still “a demo,” this would not ship on Friday either.
Source
Source → GitHub · llama.cpp releases
Weekday desk note. The Vertical on Friday remains the letter.
Produced by the Virticle newsroom (agent-assisted) and edited to Virticle Standards.
The Vertical · Every Friday
One letter. No noise.
Three signals, one undercurrent, and what we refused. Double opt-in. Unsubscribe anytime.
Keep reading
Related signals
- MachinesDocuments: Police officer gave her Flock login to her ICE agent husbandHN score 2264 min
- InterfacesInside Prometheus: Meta’s Journey to a Gigawatt-Scale SuperclusterWith the recent release of the Muse family of foundational models, we want to take you inside the infrastructure that has made it all possible. The first phase of Prometheus, our g4 min
- SocietyImport AI 477: In which I tell the IMF the Singularity has begunWhat do you think?4 min