--- format: "aidr-story-markdown/v1" id: "7932e50f9a950e28db742b8db5ff91e9a7aba49641f2323698774a5836da100a" canonical_url: "https://aidr.today/7932e50f?lang=en" title: "llama.cpp Speeds Mac Speculative Decoding 3.4x for M3 Ultra" lang: "en" requested_lang: "en" available_langs: ["en","vi"] translation_fallback: null fallback_fields: [] published_at: "2026-10-10T17:08:31.000Z" category: "Open Source" topics: ["meta","inference","open-source"] source_urls: ["https://huggingnews.com/ai/llamacpp-speeds-mac-speculative-decoding-34x-for-m3-ultra-9e14a5c5","https://x.com/qvac/status/2107043339799593421","https://x.com/qvac/status/2107924001939235207","https://x.com/SeraAndroid/status/2108963506682810506","https://marketbrief.now/ai/llamacpp-speeds-mac-speculative-decoding-34x-for-m3-ultra-9e14a5c5"] summary: "GPU optimizations merged into the llama.cpp library increase speculative decoding speed on M3 Ultra hardware from 32.1 tok/s to 110 tok/s. Metal kernels were also extended to cover all 26 weight formats supported by the Mac GPU, allowing matrix multiplies to run up to 4.4x faster." --- # llama\.cpp Speeds Mac Speculative Decoding 3\.4x for M3 Ultra > [Open the canonical story]() **Published:** 2026-10-10T17:08:31.000Z **Category:** Open Source **Topics:** meta, inference, open\-source ## Summary GPU optimizations merged into the llama\.cpp library increase speculative decoding speed on M3 Ultra hardware from 32\.1 tok/s to 110 tok/s\. Metal kernels were also extended to cover all 26 weight formats supported by the Mac GPU, allowing matrix multiplies to run up to 4\.4x faster\. ## Sources - [Story source]() - [Story source]() - [Story source]() - [Supporting source]() - [Story source]()