LiquidAI has closed the LFM2.5 family with a model that upends the dominant narrative: what matters is not just total parameters, but how many are activated for each token. The LFM2.5-24B-A2B claims 24 billion total parameters and 2.45 billion active parameters. That gap is not a technical footnote; it is the core of a strategy to bring machine reasoning to local hardware without exploding inference costs.

The spec sheet shows how. The model uses 40 layers, including 30 double-gated short convolution blocks and 10 GQA blocks, with 64 experts per MoE block and top-4 routing. In practice, each token activates only a narrow slice of the network, keeping per-token computing cost low. Context length reaches 131,072 tokens and the vocabulary includes 128,000 entries, with a stated training budget of 36 trillion tokens. Supported languages include Italian, along with English, Spanish, French, German and others. The suggested generation parameters (temperature 0.1, top_k 50, repetition_penalty 1.05) point to deterministic output, consistent with agentic use.

On performance, LiquidAI positions the model against Qwen3.6-35B-A3B, a MoE with about 50 percent more total and active parameters. On the LiveClaw benchmark, the new model scores 48.09 against 38.19. The source also reports 184 tokens per second on an Apple M5 Max and up to 21,000 output tokens per second on a single NVIDIA B300. These numbers speak to people evaluating local deployment, not just benchmark watchers.

That is the structural point. A model with 24 billion total parameters but only 2.45 billion active can shift the TCO equation for on-device agents. Fewer active parameters per token means less energy and less VRAM pressure per step. In a privacy and data sovereignty context, the ability to run a reasoning model that always thinks before it answers without sending prompts to an external cloud API has concrete implications for developers of agents handling sensitive data. The GGUF files published on Hugging Face further lower the barrier for local use.

The whole LFM2.5 family, from the 230-million-parameter model to this 24B-A2B, spans two orders of magnitude. It is a signal that the vendor is pushing a range for distributed inference, from device to single-GPU server. The challenge for local runtimes will be managing expert routing without canceling the speed gains: a theme that concerns the entire MoE ecosystem, not just this model.