Liquid AI's LFM3-1.7B Runs at 43.6 Tokens a Second on an iPhone 17 Pro, but Firms Above $10 Million Need a Paid Licence

Liquid AI released LFM3 on October 27, a family of six open-weight models of about 2 billion parameters each, built on one backbone checkpoint. The text model, LFM3-1.7B, has 1.73 billion parameters, a 256,000-token context window and fits in about 1 GB at 4-bit precision, according to its Hugging Face card.

The weights are not gated. They sit under the LFM Open License v1.0, whose exact terms the card links rather than states.

The licence line sits at $10 million

An explainer from eesel.ai describes the licence as based on Apache 2.0. Companies under $10 million in annual revenue use the weights free. Companies at $10 million or more need a paid licence quoted by Liquid sales, and revenue is counted across the whole corporate group, not one product.

Research use and qualified non-profits have no revenue limit. The licence has no copyleft, and redistribution requires attribution and a copy of the licence. That threshold comes from eesel.ai alone; the Hugging Face card does not print it.

What the card says the model is for

Liquid positions the model for local agents, tool calling, structured outputs and domain fine-tuning. The card says it is "not recommended for agentic coding or complex multi-step reasoning".

The architecture has 24 layers: 13 gated short convolutions, 5 sliding-window attention layers and 6 global attention layers. The base model saw 22 trillion tokens, followed by 1.75 trillion tokens of supervised fine-tuning and reinforcement learning stages. It supports 16 languages.

Speed on an iPhone 17 Pro, measured by Liquid

In Liquid's own tests with llama.cpp at Q4_0 and 32,000 tokens of context, the model decodes 43.6 tokens a second on an iPhone 17 Pro, using 1.90 GB of memory. MiniCPM5-2B managed 15.8 in the same test, according to eesel.ai's reading of Liquid's table. On an Apple M5 Max it reaches 237.1 tokens a second.

1. LFM3-1.7B43.6 tokens/s
2. MiniCPM5-2B15.8 tokens/s

Source: eesel.ai summary of Liquid AI's LFM3 tests, accessed 2026-10-30

Benchmarks: ahead on agents, behind on knowledge

Liquid reports an average of 49.37 across 29 benchmarks, against 47.24 for Qwen3.5-9B, a model about five times larger. All figures are vendor-reported. eesel.ai said Artificial Analysis had not yet listed LFM3.

Benchmark	LFM3-1.7B	Qwen3.5-9B
IFBench	73.03	48.84
AutomationBench	41.29	15.10
BFCL V4	71.39	62.83
SWE-bench Verified	43.40	51.20
GPQA Diamond	63.37	80.40
RULER 128K	65.56	87.11

The pattern is consistent. The small model leads on instruction following and tool calling and trails on coding, graduate-level science questions and long-context retrieval.

Gaps reported at launch

eesel.ai lists tooling gaps. The vision model's card shows llama.cpp video support still in review and vLLM support in progress. The audio and omni models have 32,000-token contexts against 256,000 for the text model.

The same write-up says early Hacker News users reported that the model answered "Oink" when asked what a cow says, with one blaming quantisation. That is a user report, not a test run here. No API price has been published, and LFM3 was not yet on OpenRouter when eesel.ai wrote.

For comparison with other small local models, a 131,000-parameter world model runs on a $5 board. On licence terms, Mistral's Large 4 weights carry a clause on hosted inference.

The Hugging Face card states no release date. The model is available in BF16, FP8, NVFP4, GGUF, MLX 4-bit and ONNX builds.