LFM2.5-2.6B: Liquid AI's On-Device Agent Model That Runs in Under 2.5 GB

Key Design Focus: On-Device Agentic Workflows

The model is purpose-built to run agentic workflows entirely on-device — planning, calling tools, web searching, and executing multi-step tasks — with data never leaving the device and zero marginal inference cost . It has native tool calling and was trained to work reliably inside real agent harnesses, including Hermes Agent, OpenClaw, and Pi .

Training Details

• Pre-training: Approximately 34 trillion tokens .
• Vocabulary: 128K tokens (doubled from prior models to better support non-Latin scripts and varied agent inputs like tool schemas and file paths) .
• Context window: 128K tokens (131,072), extended via a dedicated mid-training context-extension phase .
• Four-stage post-training pipeline :
  1. Supervised Fine-Tuning (SFT): Two rounds — broad coverage then targeted shaping on agentic tasks, tool use, reasoning, and software engineering.
  2. Teacher Specialization: One expert per domain (instruction following, math, knowledge/hallucination control, code, tool use, long context) trained via SFT + RLVR.
  3. Multi-Domain On-Policy Distillation (MOPD): Specialized teachers supervise the student model with token-level feedback during the student's own rollouts.
  4. Agentic Reinforcement Learning (Agentic RL): Multi-turn RL inside real agent harnesses using GRPO, with outcome-based rewards combining an LLM judge, programmatic checks, and a safety gate.

Memory Footprint and Decode Speeds

LFM2.5-2.6B runs in under 2.5 GB of memory . The following decode speeds are vendor-reported figures from Liquid AI:

Supported Inference Frameworks

The model supports llama.cpp (GGUF), MLX (Apple Silicon), vLLM, SGLang, ONNX, and standard Transformers .

Availability

Open-weight download on Hugging Face at LiquidAI/LFM2.5-2.6B, with separate repositories for GGUF, MLX, and ONNX quantized formats . Licensed under the LFM1.0 open-weight license .

Company Background: MIT Spinoff

Liquid AI was founded in 2023 as an MIT CSAIL spinoff by Ramin Hasani (CEO), Mathias Lechner (CTO), Alexander Amini (CSO), and Daniela Rus (MIT professor and CSAIL director) . The company's technology evolved from research on liquid neural networks at MIT .

Limitations

LFM2.5-2.6B is not recommended for agentic coding or knowledge-heavy tasks. On coding benchmarks, larger models (e.g., Qwen3.5-9B) maintain an edge, and the official Hugging Face card advises against using it for those workloads . For complex agentic coding or deep knowledge retrieval, larger models are still preferable .