Liquid AI's LFM2.5-2.6B: A 2.6B Agent That Thinks It's Bigger

Liquid AI has released LFM2.5-2.6B, a 2.6B-parameter hybrid model aimed squarely at on-device agentic workloads. The pitch is simple: you get agent-level tool use and instruction following without needing a GPU cluster. The model combines a 128K context window with agentic post-training, and the numbers are worth a second look.

Specs and Performance

The model is built on the LFM2 architecture with 30 layers (22 double-gated short convolution blocks + 8 GQA), a vocabulary of 128,000, and a training budget of 34 trillion tokens. It supports 16 languages, including Vietnamese, Thai, and Indonesian. Context length is 131,072 tokens.

Performance claims are the headline: 220 tokens per second on an Apple M5 Max and 113 tok/s on an AMD Ryzen CPU, all in under 2.5 GB of memory. That's fast enough to run real agent loops on a laptop without breaking a sweat.

Agentic Post-Training

The post-training pipeline is where Liquid AI did the interesting work. It's a four-stage process: two rounds of supervised fine-tuning, per-domain teacher specialization, multi-domain on-policy distillation, and finally agentic reinforcement learning. The RL stage trains the model inside popular agentic harnesses, exposing it to their tools, system prompts, and interaction patterns. This is a deliberate move to make the model work reliably across the agent ecosystems engineers actually use.

The model is a pure reasoning model — it always thinks before answering, adding a <think> tag in the chat template. Tool use follows a four-step protocol: define functions as JSON, call them with Pythonic syntax between <|tool_call_start|> and <|tool_call_end|> tokens, execute, then produce a final answer. You can override to JSON if you prefer.

Deployment Formats

Liquid AI ships the model in multiple formats: native (for Transformers, vLLM, SGLang), GGUF (for llama.cpp and CPU inference), ONNX (for cross-platform and edge), and MLX (for Apple Silicon). This covers the full spectrum from cloud GPU to a MacBook Air.

Supported inference frameworks include Transformers, vLLM, llama.cpp, MLX, LM Studio, and SGLang. Quick start with Transformers requires transformers>=5.0.0.

What It's Not For

Liquid AI is explicit about the model's limits: it's recommended for agentic workloads, tool use, data extraction, RAG, and long-context workflows — but not for agentic coding or knowledge-heavy tasks. That's a sensible carve-out for a 2.6B model.