Text models from the Liquid AI LFM3 series: 1.73B parameters, a tri-hybrid architecture (13 layers of gated short convolution + 5 layers of sliding window attention + 6 layers of global attention), post-trained for agent frameworks such as Hermes Agent, Pi, and OpenClaw, with a focus on running tool calls and structured output locally on smartphones and laptops.
Upvotes 178 · Likes 178 · Downloads 6311
Reasons
- Officially stated to occupy approximately 1 GB of memory at 4-bit precision; on Ryzen AI Max+ 395, Apple M5 Max, and iPhone 17 Pro, with a 32K context window, prefilling is at least 2.5 times faster and decoding is at least 1.8 times faster than all compared models, while also using the least memory.
- Official self-tests: BFCL V4 71.39 (Qwen3.5 9B scores 62.83), LiveClaw 45.60 (Qwen3.5 9B scores 35.77); however, SWE-bench Verified 43.40 remains lower than Qwen3.5 9B's 51.20, and MCP-Atlas 32.57 is also lower than Qwen3.5 4B's 41.37.
- Supports lazy loading of tools (defer_loading) combined with tool search, so not all tools need to be included in the prompt when there are many; the official team provides four single-GPU fine-tuning examples for SFT, DPO, GRPO, and continued pre-training. The LFM3-Omni-2B (which can see and hear, responding via text or voice) and LFM3-VL-2B from the same series are also on this month's leaderboard.
Usage
- Transformers: pipeline("text-generation", model="LiquidAI/LFM3-1.7B", trust_remote_code=True); the official recommendation for general tasks is temperature 0.1, top_k 50, repetition_penalty 1.1, and for mathematical tasks, temperature 1.0, top_p 0.95, top_k 20.
- Local execution: The official team provides GGUF (llama.cpp / Ollama / LM Studio), MLX 4-bit, and ONNX formats; GPU serving has FP8 and NVFP4 versions (SGLang), and vLLM is also supported.
- Thinking mode is enabled by default and can be disabled with enable_thinking=False, but the official team notes that performance is unstable when disabled; tool definitions are passed via apply_chat_template(tools=...), and the model outputs calls in XML format, with results appended as tool messages.
Risks
- The official team does not recommend it for agentic programming and complex multi-step reasoning; knowledge-intensive tasks require retrieval or search augmentation.
- The license is the LFM Open License v1.0, not Apache / MIT, so terms should be reviewed before commercial use.
- Score and speed comparisons are based on official self-tests; speed tests used a llama.cpp build with LFM3 support, and tool search on llama.cpp also requires a version compatible with LFM3 and the latest chat template.