最新模型
Liquid AI Releases LFM3, a 1.7B Text Model Using Three Hybrid Compute Methods for On-Device Agents
On October 27, Liquid AI released the open-weight LFM3 model family, built on a shared backbone for text, vision, audio, and decision-making. The text model offers a 256K context window, but the company cautions that small models are not suited to complex multi-step reasoning or agentic coding.
Liquid AI launched LFM3 on October 27. The initial release includes six models of around 2B parameters spanning text, vision, audio, multimodal tasks, and decision-making, all derived from LFM3-1.7B-Base. The technical focus is extending a small, fine-tunable shared backbone across different inputs and tasks, so developers can choose a model for their needs and build specialized on-device applications. Official announcement
The text model, LFM3-1.7B, has 1.73 billion parameters. Its 24 layers comprise 13 gated short convolution layers, 5 sliding-window attention layers, and 6 global attention layers; the local attention window is 256 tokens. This combination preserves local sequence processing while allowing information to flow across positions. The model also uses per-head gated attention and partial RoPE. Liquid AI says the design aims to control activation outliers and improve the quality-versus-speed trade-off for long contexts. Architecture overview
For training, the shared backbone used about 22T tokens. The final stage added agent trajectories and tool calls, and extended the context to 256K. Post-training for the text model combined supervised fine-tuning, mixed reinforcement learning, policy distillation from six expert teachers, and reinforcement learning on tasks run inside real agent frameworks. The model card lists tool calling, structured output, and domain fine-tuning as primary use cases; supported languages include Chinese. Model card
Liquid AI reports that its WebArena-Verified score rose from 3.02 for the previous LFM2.5-2.6B to 38.42. This is a vendor evaluation and does not establish that all websites or enterprise workflows will see similar gains. The model card also explicitly recommends retrieval for knowledge-intensive tasks and advises against using the model for agentic coding or complex multi-step reasoning. Evaluation and limitations
When deploying the model, distinguish weight size from total runtime memory: the stated size of about 1 GB at four-bit quantization should not be treated as the total requirement for long-context workloads. Ollama already supports ollama run lfm3, but the v0.40.12 release notes limit support for this model family to Apple hardware; Linux and Windows support is expected later. Engineering teams should evaluate their own Chinese-language tool workflows for parameter accuracy, long-conversation stability, latency, and peak memory before deciding whether the model can replace an existing one. Ollama model page, release notes