LFM

Liquid AI's LFM2.5 family is a serious contender for anyone building AI that runs on the device—not in the cloud. The 1.2B Instruct model punches above its weight in instruction following and tool use, rivaling much larger models in benchmarks like IFEval and BFCLv3. The new 230M variant widens the reach to IoT and wearables. What really sets LFM2.5 apart is the audio and vision models: native audio processing without pipeline latency and multilingual vision with real-world gains. Where it shines: if you need a local copilot that won't phone home, an in-car assistant that must work offline, or a privacy-first productivity tool. The open weights and permissive license (free under $10M revenue) lower the barrier. But if your use case demands a 7B+ model for complex reasoning or massive cloud-scale throughput, LFM2.5 isn't it. The 1.2B cap means limited capacity for tasks requiring deep reasoning or long-context understanding. Compared to Llama 3.2-1B or Qwen3-1.7B, LFM2.5 leads across multiple benchmarks, especially on instruction following and Japanese language tasks. Its hybrid architecture delivers faster CPU inference than typical transformer models. However, the ecosystem is smaller—fewer community LoRAs, less third-party tooling. For now, you'll rely on the provided LEAP platform and Hugging Face. In practice, the audio model's 8x speedup is real and noticeable on mobile CPUs. The detokenizer quality at INT4 is nearly lossless. But if you need streaming audio or real-time conversation, the model's latency may still be higher than cloud alternatives. Overall, LFM2.5 is a compelling pick for edge-first developers who value openness and efficiency over raw scale.