Liquid AI released LFM2.5-24B-A2B on October 13, 2026, a downloadable text model for running AI assistants on your own computer. Its expanded text capacity offers technically capable users and small-business teams more space for longer conversations and multi-step tasks.
What the larger window changes
The model’s context window—the amount of text it can work with at once—rises from 32,768 to 131,072 tokens. Tokens are the chunks into which a model divides text. Dividing the new limit by the old one gives a fourfold increase in token capacity, not a fourfold improvement in accuracy.
The larger window gives a local assistant more room for documents and task history. It does not establish that the assistant will recall every detail or complete a task correctly. Tool descriptions also use part of that space.
Unlike its predecessor, the new model always generates intermediate reasoning text before its final answer. Liquid’s documentation tells users to leave enough output capacity for that extra text.
The download is only part of the setup
This is a model to connect to assistant software, not a ready-made office assistant. Liquid’s setup guide describes connecting a local model server to Hermes Agent, OpenClaw or Pi, with compatible serving software and tool use configured.
The official GGUF repository lists several versions for local-running tools. Three illustrate how much the download size varies with the numerical precision used to store the model:
GGUF version | Download size |
|---|---|
Q4_K_M | 14.53 GB |
Q8_0 | 25.49 GB |
BF16 | 47.96 GB |
Sizes are calculated from the repository’s byte counts using 1 GB = 1 billion bytes, rounded to two decimal places. They are file sizes, not measured memory requirements.
For capacity planning, allow additional memory for context, running software and the operating system. A download that fits on disk does not establish that the full context window will fit in memory. Minimum memory and full-window speed on a reader’s machine remain unverified.
Knowledge and privacy still depend on the setup
The model card specifies text-only input: this model does not directly understand images or audio. It is separate from the Liquid vision model previously covered by RohitAI.
Liquid also advises against relying on the model alone for knowledge-heavy tasks. It recommends retrieval—bringing in information from documents or other sources—or web search.
Local processing is not an end-to-end privacy guarantee. Liquid’s tool documentation describes the surrounding application executing the model’s requests, which can involve databases and external services. The practical implication is to check where each connected tool sends information, not just where the model runs.
Commercial use has conditions
The LFM Open License v1.0 sets a $10 million annual-revenue threshold and defines the relevant legal entity to include businesses under common control. Commercial use above the threshold is not covered by this licence. The wording at exactly $10 million needs clarification from Liquid AI; a small project inside a larger company should not assume its own budget determines eligibility.
Reporting note: This AI-assisted brief draws on Liquid AI’s announcement, documentation, repository metadata and licence. RohitAI did not run the model or independently test its performance.