Liquid Claims 184 tok/s on My Mac — I Measured 174: The Speed Claim Holds
TheOyinbooke · x · 2026-10-14
The author replicated Liquid AI's official speed claim for LFM2.5-24B-A2B on a Mac Studio M5 Max: using Liquid's exact setup (llama.cpp, Q4KM, 4K input), he measured 174.1 tok/s (179.1 on short prompts), within 5% of the claimed 184. Against Qwen3-6.35B-A3B and Gemma4-26B-A4B at 4-bit in LM Studio, LFM2.5 was the fastest in the test, in both llama.cpp and MLX runtimes. The model file is 14.53GB — too big for 16GB Macs.
The catch: it's a thinking model that reasons before every reply, sometimes burning the entire 4096-token budget with empty output. Verdict: 8/10 as an agent brain on 32GB+ Macs, 5/10 for quick one-shot answers — but the quickest thinking model at making tool calls.
More from Infra
- Podcast: Marvell Analyst Day, OpenAI ARR rumors, and the 20-30 GW grid bottleneck — BenBajarin · 2026-10-14
- vLLM and Hugging Face Host Livestream on Speeding Up Local LLMs — mervenoyann · 2026-10-14
- Modal crosses $1B annualized run rate, raises $750M Series D as usage triples in 5 months — josh_wills · 2026-10-14
- Nvidia's Jensen Huang pledges $2B for Europe, $250M in compute for UK safety labs — The Decoder · 2026-10-14
- llama.cpp PR Adds Chunked GDN Kernel: Up to 18% Faster on 5090/Blackwell — jacek2023 · 2026-10-14
- WFE equipment enters structural expansion as 2nm transition obsoletes legacy tools — BenBajarin · 2026-10-14