LFM2.5-24B-A2B
An open-weight large language model released by Liquid AI using a Mixture-of-Experts architecture, with 24B total parameters and 2B activated parameters. It is designed for on-device agent scenarios and supports efficient local inference.
Core Facts
The core appeal of local AI lies in privacy, low latency, and offline availability.
Model weights cannot fill up VRAM entirely, as space must be reserved for the KV cache and inference processes; the KV cache required for a 128K context can reach as high as several GB or even exceed the model weights themselves.
Timeline (last 90 days)
LiquidAI released the new LFM2.5-24B-A2B model.
The user has not yet reached a conclusion on LFM2.5-24B-A2B's insight capability in AI-Council mode and needs longer testing.
LFM2.5-24B-A2B performs stably on mechanical tasks and has not disappointed the user in long-running recursive tasks.
On an AMD W7900 GPU using llama.cpp with the ROCm backend, the Q8 quantized version of LFM2.5-24B-A2B sustains an inference speed of about 95 tokens/second.
Data released by Liquid AI shows that LFM2.5-24B-A2B beats Qwen3.6-35B-A3B on tool-use benchmarks
For building local agent tasks (such as automated script orchestration, browser operation, file management), the speed advantage of LFM2.5-24B-A2B is more suitable; for code assistance tasks, Qwen3.6-35B-A3B is more suitable
LFM2.5-24B-A2B and Qwen3.6-35B-A3B are both open-source, locally deployable MoE architecture models, with hardware requirements within the range of high-end consumer devices
LFM2.5-24B-A2B surpassed Qwen3.6-35B-A3B on five of eight tool-use benchmarks.
LFM2.5-24B-A2B sparked discussion on Hacker News.
Activated parameters at the 2B level mean the model cannot directly compete with cloud large models in complex reasoning and long-chain planning.
25 more timeline events
All Facts (20)
The core appeal of local AI lies in privacy, low latency, and offline availability.
70%VerifiedModel weights cannot fill up VRAM entirely, as space must be reserved for the KV cache and inference processes; the KV cache required for a 128K context can reach as high as several GB or even exceed the model weights themselves.
65%UnverifiedLFM2.5-8B-A1B支持128K token的上下文窗口长度
95%UnverifiedLFM2.5-24B-A2B uses a 40-layer architecture: 30 double-gated short convolution blocks + 10 GQA attention layers, 64 experts per MoE block, top-4 routing, a 131,072 token context, and a 128K vocabulary.
60%UnverifiedThe user has not yet reached a conclusion on LFM2.5-24B-A2B's insight capability in AI-Council mode and needs longer testing.
50%UnverifiedOn an AMD W7900 GPU using llama.cpp with the ROCm backend, the Q8 quantized version of LFM2.5-24B-A2B sustains an inference speed of about 95 tokens/second.
50%UnverifiedLFM2.5-24B-A2B performs stably on mechanical tasks and has not disappointed the user in long-running recursive tasks.
50%UnverifiedData released by Liquid AI shows that LFM2.5-24B-A2B beats Qwen3.6-35B-A3B on tool-use benchmarks
50%UnverifiedLFM2.5-24B-A2B and Qwen3.6-35B-A3B are both open-source, locally deployable MoE architecture models, with hardware requirements within the range of high-end consumer devices
50%UnverifiedFor building local agent tasks (such as automated script orchestration, browser operation, file management), the speed advantage of LFM2.5-24B-A2B is more suitable; for code assistance tasks, Qwen3.6-35B-A3B is more suitable
50%UnverifiedLFM2.5-24B-A2B surpassed Qwen3.6-35B-A3B on five of eight tool-use benchmarks.
50%UnverifiedLFM2.5-24B-A2B has a total parameter count of 24B, with activated parameters at roughly the 2B level.
50%UnverifiedActivated parameters at the 2B level mean the model cannot directly compete with cloud large models in complex reasoning and long-chain planning.
50%UnverifiedRunning LFM2.5-24B-A2B on macOS + Apple Silicon via MLX (mlx-lm>=0.32.0) can achieve over 150 tokens/second within 16GB of memory.
50%UnverifiedLFM2.5-24B-A2B has 24 billion total parameters and uses an MoE architecture with 64 experts per block, Top-4 dynamic routing, and only 2.45 billion parameters activated per token.
50%UnverifiedLiquid AI officially recommends pairing the local model with a RAG retrieval system to compensate for the scale limitations of a static knowledge base.
50%UnverifiedOn Linux/Windows, the official Q4_K_M quantized checkpoint can be loaded via llama.cpp, and layer offloading (-ngl 99) enabled to start a local HTTP server compatible with the OpenAI API format.
50%UnverifiedLiquid AI's LFM2.5-24B-A2B is a mixture-of-experts (MoE) model with 24B total parameters and only 2.45B activated per token.
50%UnverifiedLFM2.5-24B-A2B is a pure-text reasoning model with no vision tower, always begins with a <think> block, and those needing image capabilities should use LFM2.5-VL-3B (the highest model in the LFM2.5 VL series).
50%UnverifiedLFM2.5-24B-A2B's KV cache grows far more slowly than a pure Transformer of the same depth, because only 10 layers are attention layers, making the memory overhead of long contexts significantly lower.
50%