Liquid AI’s 24B model targets local agent workloads

Liquid AI has released LFM2.5-24B-A2B, the final and largest model in its LFM2.5 family. The text-only mixture-of-experts model contains 24 billion parameters while activating about 2.45 billion for each token. It supports a 131,072-token context window and ships in formats for GPU servers, CPUs and Apple Silicon.

A mixture-of-experts model routes each token through a subset of its available parameters. That reduces computation during generation, although the runtime generally must still load the full weight set into memory. The active-parameter count therefore affects speed more than storage requirements.

• Total parameters: 24 billion
• Active per token: about 2.45 billion
• Context window: 131,072 tokens
• Available formats: Safetensors, FP8, GGUF, ONNX and 4-bit MLX
• Primary use: local reasoning, tool use and multi-step agent loops

Liquid reports decode speeds of 184 tokens per second on an Apple M5 Max and 114 tokens per second on an AMD Ryzen AI Max+ 395. Its published comparison shows up to a 1.7-fold speed advantage and the lowest memory use among the tested mixture-of-experts models. Quantization, prompt length, batch size and runtime configuration can materially change those results.

Forty layers, four experts per token

LFM2.5-24B-A2B retains the hybrid architecture of LFM2-24B-A2B. Its 40 layers comprise 30 double-gated short-convolution blocks and 10 grouped-query-attention blocks. Short convolutions perform efficient local sequence mixing, while grouped-query attention handles longer-range relationships with a smaller key-value cache than conventional multihead attention.

Each mixture-of-experts feed-forward layer selects four of 64 experts for a token. This sparse routing accounts for the gap between the model’s 24 billion stored parameters and its 2.45 billion active parameters.

Relative to the earlier release, the context window increased from 32,768 to 131,072 tokens, and the vocabulary expanded from about 65,000 to 128,000 tokens. Liquid performed the tokenizer expansion in place to improve the handling of non-Latin text. Base-model pretraining also increased from roughly 17 trillion to 33 trillion tokens.

Every assistant turn begins with a reasoning block, whereas the previous checkpoint did not consistently reason before answering. Applications that parse streamed output, expose raw model text or dispatch tool calls should account for that block and test how their runtime separates reasoning from final responses.