News

Releases & Models

CONFIRMED: Liquid AI launches LFM2.5-24B-A2B, an open-weight model for fast on-device agents

Liquid AI released LFM2.5-24B-A2B, an open-weight mixture-of-experts model with only 2B active parameters, built for fast private agents running on-device.

Ilustracao editorial de uma esfera de rede neural translucida com poucos nos ativos acesos, flutuando sobre silhuetas de smartphone e notebook, representando agentes de IA rapidos e privados rodando no dispositivo

The last piece of a two-order-of-magnitude family

Liquid AI closed out its LFM2.5 lineup today with LFM2.5-24B-A2B, a mixture-of-experts model with 24 billion total parameters but only 2 billion active per token. It is the largest member of a family that now spans from a 230-million-parameter version meant for wearables to this 24B flagship, all open weight and all published directly on Hugging Face and the company's own Playground the same day.

Why active parameters matter more than total size

The whole pitch of the LFM2.5 family is that most of the 24 billion parameters stay dormant for any given token, so the model runs with the speed and memory footprint closer to a 2B model while drawing on the knowledge of a much bigger one. Liquid AI built it specifically for agentic workloads that run locally: tool calling, multi-step tasks and private, on-device reasoning, with a 131,000-token context window and reasoning turned on by default, meaning it thinks through steps before answering rather than responding immediately.

What changes for builders

Both the post-trained model and its base checkpoint, LFM2.5-24B-A2B-Base, are available today in standard Transformers format plus quantized GGUF, ONNX and MLX builds, which means it already runs through llama.cpp, vLLM, SGLang and Apple's MLX stack on day one. According to Liquid AI's own benchmarks, it performs on par with Qwen3.6-35B-A3B, a considerably larger model, which is the kind of efficiency claim worth testing directly before trusting it for a production agent pipeline.

The limits worth keeping in mind

These are Liquid AI's own benchmark numbers; there is no independent third-party evaluation yet comparing LFM2.5-24B-A2B against Qwen3.6-35B-A3B or other open-weight agents under the same harness. Liquid AI is a much smaller lab than the frontier players, so community testing, long-run stability and real tool-calling accuracy will take time to show up. Running a 24B MoE locally still requires a reasonably capable GPU or Apple Silicon machine, even with quantized builds, so on-device deployment is not automatic for every device class.

MaxAssistant's read

This is a CONFIRMED launch: Liquid AI's own blog states the release plainly, and the weights are live on Hugging Face right now, which is as primary as a source gets. The open question is not whether it shipped, but whether it holds up outside Liquid AI's own benchmark suite. For teams chasing private, local agents without a frontier API bill, it is worth a real test run; for anyone who needs guaranteed top-tier reasoning today, it is a candidate to watch, not yet a default pick.

Sources

Liquid AI: LFM2.5-24B-A2B, Built for Fast On-Device Agents; Hugging Face: LiquidAI/LFM2.5-24B-A2B model card.