Editorial illustration for Liquid AI releases LFM2.5-24B-A2B, a 2B-active MoE for on-device agents
Liquid AI Releases 2B-Active MoE Model for On-Device Agents
Liquid AI has shipped LFM2.5-24B-A2B, the largest and last release in its LFM2.5 lineup, and it's built for a narrow job: running agents on hardware that's already sitting on someone's desk. The model carries 24 billion total parameters but only activates about 2 billion per token, the kind of Mixture-of-Experts design that lets a laptop or mini PC handle work that would normally need a server rack. Liquid AI trained it directly inside agent harnesses instead of tuning it for casual chat, so the target use case is multi-step tool calling, not back-and-forth conversation.
Base and post-trained weights are up on Hugging Face now, with builds for GGUF, MLX, ONNX and FP8 so it can run on machines like an Apple M5 Max or an AMD Ryzen AI Max+ 395, or get served off a single NVIDIA B300. The model handles a 131,072-token context window and reasons in 16 languages, including Arabic, Hindi and Japanese. Liquid AI's own benchmark numbers show it beating larger MoE rivals on some agentic tasks and losing ground on others, a split worth a closer look before deciding where this thing actually belongs in a workflow.
Liquid AI has released LFM2.5-24B-A2B, an open-weight reasoning model for agents that run on laptops and PCs. LFM2.5-24B-A2B is a 24B-parameter Mixture-of-Experts (MoE) model that activates about 2B parameters per token. It is the largest and final member of the LFM2.5 family.
Why this matters
For developers building agents that need to run on a laptop rather than call out to a cloud API, LFM2.5-24B-A2B is worth testing now, not later. The 2B-active design is the real story here: you get a 24B-parameter model's knowledge with inference costs closer to a 2B model, which changes the math on what's feasible for local tool-use agents on consumer hardware. Liquid AI training it inside actual agent harnesses rather than optimizing for open chat is a deliberate bet that most of the demand for this class of model is multi-step tool calling, not conversation.
The 64-expert, top-4 routing setup and the jump to 131K context after mid-training on 300B tokens suggest Liquid AI is targeting long-horizon agentic workflows specifically, not general-purpose chat benchmarks. We'd want to see independent benchmarks on actual tool-calling accuracy before treating this as a drop-in replacement for cloud-hosted agents. But for founders weighing on-device inference against API costs and latency, this is the kind of release that deserves a real evaluation, not just a glance at the parameter count.
Related Reading
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source Box CEO Aaron Levie Says AI Will Drive "Biggest Cybersecurity Upgrade
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
From Archives Researchers Warn AI Guardrails Can Be 'Completely Removed
From Archives SWE-agent: coding agents solve tasks, but setup affects wasted work
Common Questions Answered
What is the Mixture-of-Experts (MoE) design in LFM2.5-24B-A2B and how does it benefit on-device agents?
The LFM2.5-24B-A2B uses a Mixture-of-Experts architecture with 24 billion total parameters but only activates approximately 2 billion parameters per token. This design allows laptops and mini PCs to run agent workloads that would typically require server infrastructure, making local deployment feasible for consumer hardware while maintaining the knowledge capacity of a much larger model.
How does LFM2.5-24B-A2B's training approach differ from typical language model optimization?
Liquid AI trained LFM2.5-24B-A2B directly inside agent harnesses rather than optimizing it for casual chat applications. This deliberate training methodology ensures the model is specifically tuned for tool-use and agent tasks, making it more effective for developers building agents that run locally on consumer devices.
What makes LFM2.5-24B-A2B the final release in the LFM2.5 lineup?
LFM2.5-24B-A2B is the largest and final member of the LFM2.5 family, representing the culmination of Liquid AI's LFM2.5 model series. It combines the scaling capabilities and agent-specific training that defines the complete LFM2.5 product line.
Why does the 2B-active parameter design change the feasibility calculations for local agent deployment?
The 2B-active design provides inference costs comparable to a 2-billion parameter model while delivering the knowledge and reasoning capabilities of a 24-billion parameter model. This efficiency fundamentally alters the economics of running tool-use agents locally, making it practical for developers to deploy sophisticated agents on laptops and consumer PCs without cloud API calls.
Further Reading
- LFM2.5-24B-A2B: Built for Fast On-Device Agents - Liquid AI
- LFM2.5-24B-A2B - Liquid AI Docs
- LFM2-24B-A2B: Scaling Up the LFM2 Architecture - Liquid AI
- Liquid Foundation Models - Liquid AI
- Liquid Foundation Models: Complete Library - Liquid AI Docs
