Liquid AI opens LFM2.5-24B-A2B reasoning weights for local agents

Liquid AI has released the open weights for LFM2.5-24B-A2B, a reasoning-focused mixture-of-experts model designed for local assistants and multi-step agents. The October 13 release provides both a post-trained checkpoint and a separate base model on Hugging Face. Although the architecture contains about 24 billion parameters in total, Liquid AI says roughly two billion are active for each token. Active-parameter counts describe the compute path, not the amount of model weight memory a device needs to hold.

The new model builds on LFM2-24B-A2B while expanding its documented context window from 32K to 131K tokens and its vocabulary from about 65K to 128K. Liquid AI retains a hybrid design with short convolutions and grouped-query attention, alongside routed experts. Its published account describes reasoning-oriented pretraining and post-training stages that include supervised fine-tuning, teacher specialization, distillation and reinforcement learning in agent harnesses. Those details are publisher disclosures rather than independently audited training evidence.

The company reports 48.09 on LiveClaw and provides comparisons across tool use, reasoning, coding and instruction-following benchmarks. It also cites 184 output tokens per second on an Apple M5 Max and more than 21,000 output tokens per second under high concurrency on a single NVIDIA B300. Hardware settings, batch sizes and evaluation protocols matter; Open Model Weights has not reproduced these measurements. Liquid AI documents integration paths involving llama.cpp, MLX, vLLM, SGLang and ONNX, but compatibility still depends on the exact model and runtime versions.

Downloadable weights do not mean unrestricted reuse. The official repository carries the LFM Open License v1.0, whose commercial-use section conditions permission on a legal entity remaining below a US$10 million annual-revenue threshold, subject to its stated exceptions. For the evidence registry, the instruct/reasoning and base artifacts should be verified separately at pinned revisions, including file lists, license, tokenizer and generation template. The release expands local agent options without establishing that any deployed agent is independently reliable or safe.