Liquid AI’s LFM3-1.7B-Base is built to be fine-tuned, not chatted with

Liquid AI has published LFM3-1.7B-Base on Hugging Face, the pre-trained base checkpoint behind its LFM3-1.7B text model. It has 1.73 billion parameters, a 256K-token context window and was trained on 22 trillion tokens.

A base checkpoint meant for fine-tuning

The model page describes LFM3-1.7B-Base as a text-only starting point for continued pre-training and custom post-training. It is not meant for chat, tool calling or agents; the page sends those uses to the post-trained LFM3-1.7B.

The architecture is what Liquid AI calls a tri-hybrid: 24 layers made of 13 gated short convolutions, 5 sliding-window attention layers and 6 global attention layers. The vocabulary is 128,000 tokens, the weights are in BF16, and loading requires trust_remote_code=True because the model uses custom modeling code. The Hugging Face sidebar rounds the size to 2B parameters, while the spec table says 1.73B.

According to the page, pre-training ran in stages. About 15 trillion tokens of web and synthetic data came first, followed by 3 trillion tokens of code, math and science at 4K context. Mid-training added 3 trillion tokens at 32K context, then 500 billion tokens to stretch to 128K and 50 billion to reach 256K. The final 3.5 trillion tokens were curated data that included agentic traces and tool calls. Mid-training also covered chat data with and without thinking traces, so reasoning is on by default in apply_chat_template and can be switched off with enable_thinking=False.

The model covers 16 languages: English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian and Polish.

Who can use it today, and how

The weights are public under the LFM Open License v1.0. The page lists support for vLLM, SGLang and Docker Model Runner, plus fine-tuning through TRL. Liquid AI links two Colab notebooks, one for supervised fine-tuning that extracts structured JSON from messy documents and one for continued pre-training that teaches the model Turkish from raw text. Both start from LFM3-1.7B, and the page says to set model_id = "LiquidAI/LFM3-1.7B-Base" to begin from the base instead.

Liquid AI says that at 4-bit precision the model needs about 1 GB of memory, which puts phones and laptops in range. The page does not state a price; downloading is free, and the company separately runs an API console and a playground for the post-trained models.

The page lists no benchmarks for the base checkpoint, and none of its efficiency or quality statements has been independently verified. It also does not spell out what license terms apply to commercial use, so teams should read the license text before shipping a product.

A family of six built on one base

The release is part of a wider LFM3 launch. According to Liquid AI’s release blog, the first wave has six models of roughly 2B parameters, all branched from this base: LFM3-1.7B for text, three models for vision, audio and omni input, and two decision models specialized from the vision and omni variants. The company says each can be fine-tuned on a single GPU.

In the same blog, Liquid AI reports that LFM3-1.7B posts an average score of 49.37 in its own comparison table, ahead of Qwen3.5 9B at 47.24, and that it leads on instruction following and function calling. The company acknowledges that larger models still lead on agentic coding and knowledge tasks. A 1.7B model beating a model more than five times its size on an average is a vendor claim for now, and the choice of benchmarks matters.

Strategy: One base, many specialized models

Publishing the base alongside the tuned models fits Liquid AI’s pitch of specialized intelligence: customers start from one small pre-trained model and adapt it to a language, a domain or private data rather than rely on a general chatbot. The company has been moving in the same direction for weeks. Its 328M draft model for the 24B LFM2.5 mixture of experts targeted speed at the large end, with the company reporting a 2.4x average single-request gain and up to 3.1x at best. LFM3-1.7B-Base goes the opposite way, toward the smallest footprint.

Early adoption is thin: the page showed 15 downloads in the last month and 42 likes at the time of reading, and no inference provider hosts it. Those counts are a snapshot, and a base model is usually fine-tuned before anyone deploys it. The signals to watch are independent evaluations of the post-trained LFM3 models, the first community fine-tunes in languages beyond the 16 listed, and whether the 256K context holds up on small devices in practice.