On 13 October, Liquid AI released LFM2.5-24B-A2B, a reasoning model with 24 billion total parameters and roughly two billion active per token. Its weights are downloadable. The development matters to builders who want an agent to run on their own hardware instead of depending on a remote service for every model request.
Active parameters are not the file size
Sparse computation and model storage are different costs. An architecture with selectable experts uses a portion of its available parameters for an individual token. That does not make the complete set of weights equally small. Someone choosing a computer should not infer its memory requirement solely from the A2B suffix.
A practical evaluation needs separate checks for weights, working memory, context and concurrent requests. A long conversation can have a different cost from a short question using the same model. Our engineering interpretation is that local operation needs to be demonstrated with a particular configuration and workload, rather than a single parameter count.
Reported speed and comparison limits
Liquid AI reports 184 output tokens per second on an Apple M5 Max and a context window up to 131,072 tokens. It describes stronger results on some agent evaluations but weaker results on other tasks. The company also reports tightening evaluation environments after undesirable training behavior. These are vendor measurements and explanations, not independently established universal throughput.
A useful internal comparison would keep instructions, tools and completion criteria constant. An agent that responds quickly but needs three corrections may not finish the job sooner. First-token latency does not describe the complete process. For an application, time to a verified result, failures and sustained hardware consumption are also meaningful.
Connecting a local agent
The official guide describes a server with an OpenAI-compatible interface and a separate agent harness connecting to it. Tool calling needs the appropriate configuration. The guide explicitly treats the larger model as an option with greater memory requirements. Having model weights and having a functioning execution loop are separate achievements.
Consider an assistant organizing internal documents. Its model could run locally while a search tool sends content to an external service. Privacy would then depend on the route taken by each action. This is our hypothetical example, not an identified LFM defect: permissions, outside services and the results log must be inspected.
What the license permits
The model page identifies LFM Open License 1.0. The published license sets a $10 million annual-revenue threshold for free commercial use and requires preservation of licensing notices when distributing the model. Downloadable weights consequently do not mean an unconditional Apache 2.0 license. A proposed use must be assessed against the full terms.
For a small team, the main question is whether control over the environment justifies configuration and maintenance. Larger organizations also need to check the licensing scope. The release expands choices for privately operated infrastructure, while leaving memory, quality and data-access evaluation as essential responsibilities.
Our proposed evaluation would begin with a bounded task and sample documents whose correct result is known. The team could record which files were read, which actions were proposed and whether an approval was required before changing anything. A saved output alone would be weaker evidence than a reproducible run with an inspected action history.
Sources were read on 14 October 2026 between 09:25 and 09:35 Europe/Sofia. Assessments and proposed checks are editorial analysis; SIExpo has not independently tested the product. The cover is an AI illustration, not a photograph of an actual event.
