Quick Facts
- Liquid AI released LFM2.5-2.6B on August 4, 2026, a 2.6B-parameter model built for agentic workloads on local hardware.
- The model hits 220 tokens/s on Apple M5 Max and 30 tokens/s on a smartphone, all under 2.5 GB of memory.
- Both the base and post-trained checkpoints are available as open weights on Hugging Face.
Liquid AI released LFM2.5-2.6B on August 4, 2026, a model designed to run multi-step AI agents entirely on local hardware, from smartphones and laptops down to a Raspberry Pi 5.
The model has 2.69 billion parameters, a 128,000-token vocabulary, and a 131,072-token context window. It supports native tool calling and covers 16 languages. Pre-training used approximately 34 trillion tokens.
A Different Architecture
LFM2.5-2.6B is built on Liquid’s LFM2 hybrid architecture, which pairs gated short convolution blocks with grouped-query attention blocks. Attention accounts for only 37% of the model’s layers. The remaining 63% are convolution blocks.
Standard transformer models use softmax attention at every layer, which scales quadratically with sequence length and demands large KV caches. The convolution blocks in LFM2 maintain a fixed-size state per token, sidestepping that memory problem entirely. The result is a model that runs under 2.5 GB of memory across CPU hardware.
Speed and Benchmark Results
The model reaches 220 tokens per second on an Apple M5 Max, 113 tokens per second on an AMD Ryzen CPU, and 30 tokens per second on a smartphone. At high concurrency on a single H100, it approaches 15,000 output tokens per second.
On ToolSandbox, the primary benchmark for agentic task performance, LFM2.5-2.6B scored 77.83 against 76.44 for Qwen3.5-9B, a model roughly 3.6 times larger. In a separate test by AI client platform Atomic Chat, LFM2.5-2.6B completed 35 tool calls across three tasks 3.7 times faster than DeepSeek-V4-Flash, which carries 284 billion parameters.
The model trails Qwen3.5-9B on math and code benchmarks. It scored 51.87 on AIME25 versus 56.07 for Qwen3.5-9B, and 59.41 on LiveCodeBench versus 69.86.
Who This Is For
Maxime Labonne, Liquid AI’s Head of Post-Training, told VentureBeat the model targets use cases where cloud inference is not an option. “I do also believe that the best models will be in the cloud, and there’s no problem with that,” he said. “We want to make models for another type of user, and the best way of describing it is: you should use [edge AI] when you can’t use a cloud model.”
Regulated industries, air-gapped environments, and applications handling sensitive data are the clearest fits. Because inference runs on-device, data never leaves local hardware and the per-run cost is effectively zero.
Labonne said the team designed the architecture around CPU performance from the start, rather than optimizing for GPU benchmarks. “What we want to show is that it’s a really good trade-off, because you get the level of quality that you get with much bigger models, but in a tiny, tiny form factor,” he said.
Agentic Design and Fine-Tuning
Liquid built LFM2.5-2.6B with agent frameworks in mind, not conversational interfaces. “Models are not consumed in chatbots anymore. They’re really consumed through agentic harnesses,” Labonne said.
For production deployments, Labonne recommends fine-tuning. “If you fine-tune it well, it’s going to match the performance of GPT and Claude,” he said, noting that the barrier to fine-tuning has dropped considerably. Both the post-trained checkpoint and a base checkpoint for developers are available now on Hugging Face.
This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.