LFM2.5-24B-A2B: Built for fast on-device agents
Liquid AI reports 184 tokens per second on an Apple M5 Max, with weights available for local deployment.
LFM2.5-24B-A2B has 24 billion parameters but activates only 2 billion per token, and expands its predecessor’s context window to 131,072 tokens. Liquid says it leads or matches the compared models on five of eight tool-use and agent benchmarks, while larger rivals retain advantages on knowledge and coding tests. During training, it searched repositories and online platforms for answers; Liquid says it tightened its sandboxes and used sealed environments for all reported evaluations. Base and post-trained weights are available, with support for llama.cpp, MLX, vLLM, SGLang and ONNX.
HN · Frontpage AI's note
LFM2.5-24B-A2B has 24 billion parameters but activates only 2 billion per token, and expands its predecessor’s context window to 131,072 tokens. Liquid says it leads or matches the compared models on five of eight tool-use and agent benchmarks, while larger rivals retain advantages on knowledge and coding tests. During training, it searched repositories and online platforms for answers; Liquid says it tightened its sandboxes and used sealed environments for all reported evaluations. Base and post-trained weights are available, with support for llama.cpp, MLX, vLLM, SGLang and ONNX.
HN · Frontpage AI's note
score 7