Model releases · 11 August 2026
Liquid AI's largest LFM2.5 model ships with more distribution formats than most releases three times its size, and a licence that is not Apache.
| Released by | Liquid AI |
|---|---|
| Published | 6–7 August 2026 |
| Parameters | 2.6B |
| Architecture | Hybrid, built on LFM2 (Lfm2ForCausalLM) |
| Context length | 128,000 tokens |
| Modality | Text only |
| Post-training | Tuned for agentic workloads |
| Licence | LFM Open License v1.0 — not Apache or MIT |
| Weight format | safetensors — 5.41 GB |
| Official builds | GGUF, MLX, MLX-NVFP4, MLX-MXFP4 |
| Hosted at | LiquidAI/LFM2.5-2.6B on Hugging Face |
LFM2.5-2.6B is the largest member of Liquid AI's LFM2.5 family, a line designed from the start for on-device deployment rather than scaled down from a server model. It builds on the LFM2 architecture with a 128K context window and post-training aimed specifically at agentic workloads.
Liquid describes it as general-purpose and text-only. Smaller siblings in the family cover 1.2B and 350M.
Most labs at this size publish safetensors and leave conversion to the community. Liquid publishes GGUF, MLX, and two four-bit MLX variants — NVFP4 and MXFP4 — from its own account, alongside a LeapBundles repository for its deployment tooling.
For anyone shipping on Apple hardware, official MLX builds remove a step that usually involves trusting a third-party conversion.
The model card declares lfm1.0, Liquid AI's own open licence, rather than Apache 2.0 or MIT. That distinguishes it from most models in this size class and is worth reading before commercial deployment, particularly for anyone redistributing the weights inside an application.
Liquid also published LFM2-1.2B-Longevity and LFM2-2.6B-Longevity in the same window — domain-adapted variants rather than general releases, which suggests the LFM2 base is being used as a platform for vertical models.
| Runtime | Supported | Notes |
|---|---|---|
| llama.cpp | Yes | Official GGUF published by Liquid AI. The lfm2 architecture has upstream support. |
| MLX | Yes | Official MLX builds, including NVFP4 and MXFP4 four-bit variants. |
| transformers | Yes | Reference implementation from the lab. |
| Vision input | No | Text only. |
A 2.6-billion-parameter hybrid model from Liquid AI designed for on-device deployment, with a 128K context window and post-training aimed at agentic workloads.
Liquid AI's own LFM Open License version 1.0, not Apache 2.0 or MIT. Read it before redistributing the weights inside an application.
Yes. GGUF, MLX, MLX-NVFP4 and MLX-MXFP4 are all published by the lab rather than left to the community.
No. It is a text-only model.
128,000 tokens.
OnDevice LLM is a private AI assistant that runs entirely on your iPhone — no account, no cloud, and nothing you type leaves the device.
Published 11 August 2026