AIToolScan

LFM2.5 1.2B Instruct Overview

LFM2.5 1.2B Instruct is a general-purpose instruction-tuned language model from Liquid AI, part of the Liquid Foundation Models 2.5 (LFM2.5) family built for on-device deployment. With only 1.17 billion parameters, it rivals much larger models in quality while running fast and privately on laptops, phones, and edge hardware.

• Best-in-Class Performance: Outperforms other sub-2B models on GPQA (38.89), MMLU-Pro (44.35), IFEval (86.23), Multi-IF (60.98), and AIME25 (14.00), beating Llama 3.2 1B Instruct and Gemma 3 1B IT while competing with Qwen3-1.7B.
• Ultra-Fast On-Device Inference: 239 tokens/sec decode on AMD CPU and 82 tokens/sec on mobile NPU, running under 1GB of memory with no cloud needed.
• Hybrid Liquid Architecture: 16 layers of 10 double-gated convolution blocks plus 6 grouped-query attention (GQA) blocks cut memory traffic and KV-cache overhead for fast prefill and decode.
• Scaled 28T-Token Training: Extended pre-training from 10T to 28 trillion tokens with large-scale multi-stage reinforcement learning for reliable instruction following.
• 32K Context Window: Native 32,768-token context ideal for agentic tasks, data extraction, and retrieval-augmented generation (RAG).
• 8-Language Support: English, Arabic, Chinese, French, German, Japanese, Korean, and Spanish.
• Native Tool Use: Function calling with Pythonic tool_call tokens for agentic workflows, plus a ChatML-like chat template.
• Day-One Ecosystem: Supported at launch by llama.cpp (GGUF), MLX for Apple Silicon, vLLM, ONNX Runtime, Transformers, and LM Studio.
• Open License: Released under the LFM Open License 1.0 (Apache 2.0-based) — free for individuals and companies under $10M annual revenue with no copyleft obligations.