How much does LFM2.5-2.6B cost to use?

LFM2.5-2.6B is free to download and run under Liquid AI's LFM Open License 1.0, which is based on Apache 2.0. Companies with annual revenue under $10M USD may use it commercially at no cost; larger companies must contact Liquid AI for a commercial license. There is no hosted API tier.

How does LFM2.5-2.6B compare to Gemma-4 and Qwen3.5?

LFM2.5-2.6B leads every instruction-following benchmark Liquid AI tested and beats Gemma-4 5.1B and Qwen3.5 4.7B on tool use, trailing only Qwen3.5-9B on one BFCLv4 subtest. It lags Qwen3.5-9B on math (AIME25 51.87 vs 56.07) and code (LiveCodeBench 59.41 vs 69.86) — Qwen3.5-9B is 3.6x larger.

What hardware can run LFM2.5-2.6B?

LFM2.5-2.6B fits in under 2.5 GB and runs at 30 tok/s on modern phones, 113 tok/s on an AMD Ryzen AI Max+, 220 tok/s on an Apple M5 Max CPU, and ~15K tok/s at high concurrency on an H100 GPU. Data never leaves the device.

Which inference frameworks support it?

LFM2.5-2.6B ships with day-zero support for llama.cpp (GGUF), Apple MLX, vLLM, SGLang, and ONNX. Base and instruction-tuned checkpoints are on Hugging Face under LiquidAI/LFM2.5-2.6B, and a WebGPU browser demo runs entirely client-side.

Liquid AI · 2026-08-04 · major

LFM2.5-2.6B — Liquid AI's 2.6B on-device agent competes with 4x-larger models

Liquid AI released LFM2.5-2.6B, an open-weight 2.6B agent model trained on ~34T tokens with a 128K context. It runs 30 tok/s on a phone and 220 tok/s on Apple M5 Max, and leads Gemma-4 and Qwen3.5 on tool use and instruction following.

Liquid AI LFM2.5-2.6B model card social thumbnail on Hugging Face

A 2.6B open-weight agent that plans, calls tools, and runs entirely on phones, laptops, and robots.

Key specs

Parameters	2.6B
Context window	128K
Speed (m5 max)	220 tok/s

Quick facts

Maker	Liquid AI
Parameters	2.69B (dense, 30 layers)
Architecture	22 double-gated conv + 8 GQA layers
Context window	131,072 tokens (128K)
Training data	~34T tokens, 16 languages
License	LFM Open License 1.0 (Apache-2.0 based, revenue cap $10M)
Availability	Open weights on Hugging Face; llama.cpp, MLX, vLLM, SGLang, ONNX

What is it?

LFM2.5-2.6B is Liquid AI's new 2.6B-parameter open-weight language model, built for agentic workflows that run entirely on-device. It ships with base and instruction-tuned checkpoints on Hugging Face, a 128K context window, and support for 16 languages.

How does it work?

The 30-layer hybrid stacks 22 double-gated convolution blocks with 8 grouped-query attention layers, pre-trained on ~34T tokens. Post-training turns the base into an agent through supervised fine-tuning, per-domain teacher distillation, multi-domain on-policy distillation, and a final agentic reinforcement-learning stage.

Why does it matter?

LFM2.5-2.6B lets developers deploy planning, tool calling, and multi-step reasoning without a cloud round-trip — 30 tok/s on a phone and 220 tok/s on an M5 Max CPU — while beating Gemma-4 and Qwen3.5 4.7B on tool use and instruction following. Data stays on the device and the marginal cost per run is essentially zero.

Who is it for?

developers shipping local agents on phones, laptops, or robots

Frequently asked questions

How much does LFM2.5-2.6B cost to use?: LFM2.5-2.6B is free to download and run under Liquid AI's LFM Open License 1.0, which is based on Apache 2.0. Companies with annual revenue under $10M USD may use it commercially at no cost; larger companies must contact Liquid AI for a commercial license. There is no hosted API tier.
How does LFM2.5-2.6B compare to Gemma-4 and Qwen3.5?: LFM2.5-2.6B leads every instruction-following benchmark Liquid AI tested and beats Gemma-4 5.1B and Qwen3.5 4.7B on tool use, trailing only Qwen3.5-9B on one BFCLv4 subtest. It lags Qwen3.5-9B on math (AIME25 51.87 vs 56.07) and code (LiveCodeBench 59.41 vs 69.86) — Qwen3.5-9B is 3.6x larger.
What hardware can run LFM2.5-2.6B?: LFM2.5-2.6B fits in under 2.5 GB and runs at 30 tok/s on modern phones, 113 tok/s on an AMD Ryzen AI Max+, 220 tok/s on an Apple M5 Max CPU, and ~15K tok/s at high concurrency on an H100 GPU. Data never leaves the device.
Which inference frameworks support it?: LFM2.5-2.6B ships with day-zero support for llama.cpp (GGUF), Apple MLX, vLLM, SGLang, and ONNX. Base and instruction-tuned checkpoints are on Hugging Face under LiquidAI/LFM2.5-2.6B, and a WebGPU browser demo runs entirely client-side.

Try it

huggingface.co/LiquidAI/LFM2.5-2.6B