Liquid AI

Overview

Liquid AI publishes the LFM (Liquid Foundation Model) series — small open-weight models built to run entirely on-device, on phones, laptops, PCs and robots, rather than behind a hosted API (source).

The positioning is a deliberate inversion of the frontier-lab trade: instead of competing on capability at any parameter count, the LFM line competes on capability per gigabyte of device memory, with the stated consequences that "data never leaves the device" and "the marginal cost of each run is essentially zero" (source).

This page was created on 2026-08-05, on the release of LFM2.5-2.6B. Facts here are limited to what that announcement and its coverage stated; company history, funding and headcount were not established on the run that created it.

Key People

unknown — no named individual appeared in any source read on the run that created this page.

Models & Products

• LFM2.5-2.6B — 2.69B on-device agentic model, released 2026-08-04
• The wider LFM2 / LFM2.5 family, published on Hugging Face under the LiquidAI organization: sibling checkpoints named in the sources include LFM2.5-350M, LFM2-1.2B and LFM2-2.6B (source)

Recent Activity

• 2026-08-20 / 2026-08-19: Two inference-efficiency releases in two days, neither of which is a new model — LFM2.5-DSpark adds speculative decoding to the LFM2.5 family via ~300M-parameter draft models: the draft proposes a block of nine candidate tokens and the target verifies the whole block in a single forward pass. Vendor-stated up to 3.18× on an H100 and up to 2.87× on an M4 Max MacBook Pro, with outputs unchanged — this is decoding speed, not quality — for a stated small memory increase. Targets named: LFM2.5-1.2B-Instruct, LFM2.5-2.6B and LFM2.5-8B-A1B. One day earlier, Q4_0 GGUF checkpoints trained with Quantization-Aware Distillation (a high-precision teacher distilled into a quantized student) for LFM2.5-230M, 350M, 1.2B-Instruct and 2.6B, reported to recover up to 97% of BF16 accuracy. Why it matters: this lab's whole positioning is memory footprint and decode speed on device, and both releases buy those without touching the model's capability — the same freeze-the-weights-move-everything-else pattern the papers in Post-Training Scaling are running at the training end. Every figure is vendor-stated: no benchmark is named for the 97%, no batch size or sequence length for the speedups, and no draft-block acceptance rate, which is what a speculative-decoding speedup actually depends on. LFM2.5-8B-A1B is named here and this wiki holds no page for it — nothing read gives its parameters, context window, licence or release date (source)

• 2026-08-04: LFM2.5-2.6B released — an agentic model that runs entirely on-device. 2.69B parameters in under 2.5 GB of memory, pre-trained on ~34T tokens, 128K context after a mid-training extension phase. Decode speeds published for three hardware classes: 220 tokens/s on an M5 Max, 113 tokens/s on a Ryzen AI Max+, ~30 tokens/s on phone-class hardware. Post-training ran four stages including Agentic RL inside live harnesses (OpenClaw, Hermes Agent) rather than synthetic traces or offline distillation alone. Base and post-trained checkpoints plus a GGUF build published on Hugging Face, with day-one support in llama.cpp, MLX, vLLM, SGLang and ONNX. The licence is not recorded here — sibling LFM2 checkpoints carry an lfm1.0 tag, but that was not readable for this checkpoint and a family tag is not evidence about a specific release. → LFM2.5-2.6B (source) (Hugging Face blog) (Liquid AI)

Strategic Position

Occupies a lane the open-weight releases this wiki tracks have mostly left empty. The Western open-weight frontier recorded here runs large: Inkling at 975B/41B, Laguna S 2.1 at 118B/8B — both open, both far past what a phone holds. The Chinese open-weight releases run larger still (Kimi K3 at 2.8T, GLM-5.2 at 744B). LFM2.5-2.6B is two to three orders of magnitude below any of them and is not competing with them; it is competing with not running a model at all on the device in question.

The nearest comparison in this wiki is Gemma 4 12B — Google DeepMind's 12B open-weight multimodal model positioned at "16GB laptop" — which is roughly 4× larger by parameter count and targets laptops rather than phones.

The claim that distinguishes this release from earlier small models is agentic: tool calling and multi-step task execution at 2.6B, trained by RL in live agent harnesses. Whether that holds up is not settled here — no independent measurement of this checkpoint was found (source).

Related

• Agents (LLM Agents)
• Agentic Reinforcement Learning
• Open-Weights Policy Fight
• LFM2.5-2.6B
• Gemma 4 12B