HeyHomees Lab Notes · No. 3 · Liquid AI · Published October 16, 2026
By HeyHomees Newsroom ·
In 30 seconds
- A 24B-parameter model with just 2.45B active parameters per token for cheap reasoning.
- Runs at 184 tokens/second on an Apple M5 Max and up to 21K tokens/second on a NVIDIA B300 GPU.
- Scores 48.09 on LiveClaw, beating a 35B-parameter competitor that scores 38.19.
What is it?
LFM2.5-24B-A2B is a 24B-parameter text-only reasoning model that activates only 2.45B parameters per token. It is designed for agentic workloads, tool calling, and on-device inference. The model includes 40 layers, 64 experts per MoE block, and a context window of 131,072 tokens.
How does it work?
The model uses a mixture-of-experts architecture with top-4 routing, activating a small subset of experts for each token. It always generates a chain-of-thought before the final answer, enabling cheap reasoning tokens. Training involved 36 trillion tokens, with pre-training, mid-training, and multiple agentic fine-tuning stages.
What does it need to run?
The model can be served with Transformers, vLLM, SGLang, llama.cpp, MLX, and ONNX Runtime. Hardware examples include Apple M5 Max for CPU inference and NVIDIA B300 GPU for high-throughput GPU serving.
By the numbers
- Total parameters: 24.0B
- Active parameters: 2.45B
- Context length: 131,072 tokens
- Apple M5 Max inference speed: 184 tok/s
- NVIDIA B300 GPU inference speed: 21K output tok/s
- LiveClaw score: 48.09
The honest read
The reported LiveClaw score of 48.09 comes from LiquidAI’s own evaluation and shows the model is competitive with larger MoE models, but the benchmark set is limited to tool-use and agentic tasks. Inference speed figures (184 tok/s on Apple M5 Max, 21K tok/s on NVIDIA B300) are measured on specific hardware and may not reflect performance on other devices. The model’s active-parameter count (2.45B) reduces per-token compute, yet total parameter count remains 24B, so memory requirements stay high for full-model loading. The release does not provide external third-party validation of these numbers, and the model is described as “not recommended for knowledge-heavy tasks on its own,” indicating limitations for pure QA without retrieval augmentation.
What we don't know yet
- Exact memory footprint (GB) for loading the full 24B model.
- Performance on non-agentic benchmarks such as pure knowledge retrieval.
- Availability of quantized or compressed variants beyond the listed formats.
Sources
- Model card, Hugging Face (primary)
- Hugging Face Blog
Researched and drafted with AI from the sources above; every figure is checked against them by code before it is published, and a line that fails is dropped.