A local agent must fit both the device and the task. Fast token generation helps, but a system that needs more attempts or longer reasoning can still take more time to finish useful work. Privacy and responsiveness depend on the complete application, including its tools.
Liquid AI's October 13 release of LFM2.5-24B-A2B describes a reasoning model with 24 billion total parameters and two billion active parameters per token. The company reports support across local and server inference runtimes, together with benchmark and hardware measurements. These are vendor-reported results under specified configurations. Primary source.
Sparse compute is not total footprint
Active parameter count describes only part of deployment cost. The full model weights, context state and serving software still consume memory. Hardware-specific throughput also changes with precision, prompt length, batch size and concurrency.
The company's own comparison shows a mixed capability profile, with stronger results on some tool-use and instruction-following tests and weaker results on several knowledge and coding measures. That variation makes task selection more informative than a single overall ranking.
For a local workflow, tool calls create another boundary. Running the model on the device does not keep information local if the application sends it to an external search or business service. The data path must be assessed independently of model placement.
Count completed work
Near term, developers can compare the model on a representative set of tasks, measuring completion, retries, memory, latency and total generated tokens. A useful benchmark includes unsuccessful cases and the fallback to a larger model or a person.
Local reasoning models could widen the range of continuously available assistants if these costs remain manageable. The next evidence is sustained application performance on the intended hardware, with enough context and tool use to resemble real work. Tokens per second alone cannot establish that outcome.
Source organization: Liquid AI. Source statements are attributed; operational implications and suggested evaluation criteria are editorial analysis.
References
Questions, news tips or corrections? Contact [email protected].
