LFM2.5-24B-A2B LFM2.5-24B-A2B is Liquid AI's largest model in the LFM2.5 family, a 24B-parameter Mixture-of-Experts reasoning model with only 2B active parameters per token. Designed for efficient local inference on laptops and PCs, it supports 131K context, agent harnesses like OpenClaw and Hermes, and runs at high speeds such as 184 tokens/s on Apple M5 Max. Open weights available on Hugging Face with support for llama.cpp, MLX, vLLM and more.
Visit productFeatured media
Loading media…
View on XCosigns Received
“Honestly, I didn't expect LFM2.5-24B-A2B to be this good at AWS DevOps tasks. Tested it on my college AWS lab assignment, and it managed to complete the entire EBS task on its own and score full marks. The whole process took around 20 minutes, and I forgot to record the full thing. The funny part? My Liquid AI API credits are almost gone from all this testing. 😭”
“congrats on the release, this one’s great! got it running on my RTX 3060 with full 131K context at 87 tok/s 🔥”
“¡OJO! A los que viven esperando un nuevo Qwen similar al 35B/A3B: Estoy probando pi 1.2.0 en local con el modelo LFM2.5 24B/A2B Q3_K_XL (Unsloth) y funciona muy bien con el último llama.cpp. Es muy rápido para tareas agénticas.”
“Great, at last I have proper model for my MacBook 32GB M1 Max. The speed is blazing very fast even for about 110k context window, about 90toks/sec decode and 1000toks/sec prefill. Much more than enough. For short test, quality is good, at least for my Hermes agent, there were not so many loop/think as Qwen 3.8. Qwen is good but too slow for my MacBook. Enough room now for something else like local cloud. Thank you @liquidai”
“Got Liquid AI's newly released LFM2.5-24B-A2B running locally: - 276 tok/s single-user (247 tok/s at 32k context) on a single RTX PRO 6000 Blackwell - ~69,000 tok/s prompt processing with the official FP8 checkpoint on vLLM - went 25/25 on my tool-calling checks.”
“My first Impression of LFM2.5 24B A2B is that it's over 300t/s in any gpu and you can have as much context you want.”
“LiquidAI just dropped LFM2.5-24B-A2B, running it on a single RTX 3060. 12GB VRAM, 16GB RAM. (config below) full 131K context. ~87 tok/s decode, ~1064 tok/s prefill. 94% VRAM (11.6/12.3 GB), the rest sits in system RAM. 24B total, 2.45B active, and only 10 of its 40 layers are attention, so long context barely costs VRAM. running Liquid’s QAD Q4_0 build, trained to win back the quant loss. it was RL’d inside Hermes, Pi, OpenClaw and OpenCode, the harnesses people actually run local agents in. everyone wanted a Qwen3.8-35B-A3B and it never came. this might be the one to fill that slot. running my agentic coding benchmark on it with Pi next! config: llama-server -m LFM2.5-24B-A2B-QAD-Q4_0.gguf -ngl 99 --n-cpu-moe 12 -c 131072 -fa on --jinja -np 1 --cache-type-k q8_0 --cache-type-v q8_0 --temp 0.1 --top-k 50 --top-p 1.0 --min-p 0.0 --repeat-penalty 1.05 --reasoning on”