Models & Labs
llama.cpp v0.7.0: MoE caching and speculative decoding advances
llama.cpp Releaseshigh confidence
Why it matters
- →Host-memory expert caching significantly reduces latency for MoE models by keeping offloaded parameters accessible via LRU.
- →LiliCorr and GLM5-Next speculative decoding expand the range of architectures that benefit from draft-based speedups.
- →Expanded support for decision and multimodal models positions llama.cpp as a more versatile inference runtime.
llama.cpp v0.7.0 introduces significant performance optimizations for Mixture of Experts (MoE) models via a new host-memory expert cache, which can accelerate decode speeds by up to 2.2x. The update adds support for LiliCorr and GLM5-Next speculative decoding techniques, improving efficiency for draft-based inference. A comprehensive list of new model architectures is now supported, including K2 Horizon, Prism Bonsai 2, EmbeddingGemma-2, and various decision models from LiquidAI and vLLM. The server component features a refactored chat session API and changes its default listening port to 9931.
Read originalThe story around this
Earlier coverage that leads up to this article, and what followed. Lines connect each piece to the closest one after it, converging here.
How we got here
- 1Llama.cpp adds GLM-5.2 speculative decoding support
llama.cpp Releases · July 30, 2026 · Same story
- 2llama.cpp b11339 fixes cache reallocation crash
llama.cpp Releases · October 2, 2026 · Same story
Follow this story
Open the full story →llama.cpp v0.6.0 adds decision model API and Apple Silicon speedups
3 developments
- Oct 6 · llama.cpp Releases
llama.cpp v0.6.0 adds decision model API and Apple Silicon speedups - Oct 10 · llama.cpp Releases
llama.cpp UI adds offline context filtering and favorites↳ llama.cpp UI adds offline context filtering and persistent favorite marking for local model management - Oct 16 · llama.cpp Releases
llama.cpp v0.7.0: MoE caching and speculative decoding advances (This article)↳ llama.cpp v0.7.0 adds MoE caching for 2.2x speedups and LiliCorr speculative decoding support
More from llama.cpp Releases
Coding Toolscoding
llama.cpp b12022 adds ROCm 10 and Snapdragon support
This release quietly cements llama.cpp as the universal inference runtime by finally bringing ROCm 10.0 to Linux and Windows alongside CUDA 13.4, effectively closing the hardware gap for AMD users who were previously left behind. The inclusion of native Snapdragon support on Linux marks a significant step toward running local LLMs on ARM-based mobile chips without emulation overhead. While KleidiAI on Apple Silicon is disabled in this build, the broader expansion to diverse accelerators ensures that developers aren't locked into NVIDIA ecosystems anymore.
llama.cpp Releases
Coding Toolscoding
llama.cpp b12023 fixes shared prefix bug
This release addresses a specific inference artifact where decision prefixes were incorrectly bleeding into output embeddings. By capping shared prefixes at the first output position, llama.cpp ensures that option tokens remain distinct for parent and child tasks. This is a targeted fix for multi-step reasoning or voting workflows rather than a performance boost. It stabilizes complex generation patterns without changing the underlying model architecture.
llama.cpp Releases
Coding Toolscoding
llama.cpp b12024 adds Snapdragon and ROCm 10
This release quietly expands llama.cpp's hardware support to include Qualcomm Snapdragon devices with CPU, Adreno GPU, and Hexagon NPU acceleration. It also updates ROCm to version 10.0 for both Linux and Windows, keeping AMD users in step with the latest driver stack. CUDA builds now default to 13.4 libraries alongside 12.x options, reflecting the industry's gradual shift toward newer NVIDIA toolkits. While no new model architectures are introduced, the binary matrix grows significantly, ensuring local inference remains viable on a wider array of edge and server hardware.
llama.cpp Releases
More in Models & Labs
© The Verge AIModels & Labsother
Apple Intelligence arrives in China via iOS 27.3
Apple is finally unlocking its AI suite for Chinese users with the iOS 27.3 beta, marking a significant regulatory milestone after months of approval delays. The rollout relies on Alibaba’s Qwen model rather than OpenAI, integrating features like notification summaries and visual intelligence directly into the OS. By sidestepping the new Siri in favor of established text and image tools, Apple is prioritizing immediate utility over the full generative experience. This move effectively brings Apple’s AI capabilities to a massive new market, albeit with localized constraints that distinguish it from the global release.
The Verge AI
© Matt WolfeModels & Labsmodels
Google Releases Gemini 4 Carbon for Cost-Efficient Inference
Google launched Gemini 4 Carbon, a new model variant aimed at providing high performance at a lower cost for inference tasks.
Matt Wolfe
© Lev SelectorModels & Labsmodels
Google Gemini 4 Carbon Leads LM Arena Leaderboard
Google's new model, Gemini 4 Carbon, has taken the top spot on the LMSYS Chatbot Arena leaderboard.
Lev Selector