Geode · LFM3-1.7B for phones

A model pack for Geode, an Android app that runs language models entirely on the phone: no cloud, no account. LFM3-1.7B is Liquid AI's small text model: a hybrid of short convolutions and attention (13 convolution layers, 11 attention layers), with thinking and tool calls. Geode runs it with its own engine.

Packs

Folder Weights Size Quality (KL vs BF16 / top-1, on the model's own answers) For
Q4/ every matrix 4-bit (GPTQ, the output head too), the embedding 8-bit; norms, convolution kernels and attention gates exact 1.26 GB (1.17 GiB), about 1.1 GB in memory 0.037 / 93.1% any 64-bit phone with 4 GB of RAM or more

Speed on a Snapdragon 8 Elite phone (RedMagic 10 Pro): a 1,968-token prompt reads at ~2,150-2,210 tokens/s (first token after ~0.9 s) on the NPU - whole layers, attention and the convolutions included, on Hexagon's matrix units, from the same copy of the weights the processor reads - and the answer is written on the processor at ~29-32 tokens/s. Without the NPU the processor reads a prompt at ~100 tokens/s.

LFM3 thinks in English and answers best in English; asked in another language it answers in it, more weakly - a system line naming the language helps.

The packs are Geode's own format (geode.json describes the files). They are not GGUF and do not load in other runtimes.

What we changed

This pack is a modified, converted form of Liquid AI's LFM3-1.7B, not the original files:

  • the weights re-quantized from the BF16 checkpoint: every linear matrix to 4 bits with GPTQ (a scale per 16 values, calibrated on ~200K tokens of chat, tool-call, code and multilingual text), the token embedding to 8 bits; the norms, the convolution kernels and the attention output gates kept exact (FP32);
  • converted to Geode's pack format: dense.bin (the layers and the output head), embed.bin (the token embedding), tokenizer.bin (the tokenizer in Geode's binary form), geode.json (the configuration and every file's size and SHA-256), LICENSE (the LFM Open License v1.0, unchanged).

Use

  1. Install Geode (Google Play beta: geode.tifu.xyz).
  2. Models → Download from Hugging Face → Maximilian228/Geode-LFM3-1.7B.

Requirements: Android 11+, a 64-bit ARM processor with the dot-product instructions (ARMv8.2), and free storage for the pack.

License and attribution

LFM3-1.7B is made by Liquid AI and licensed under the LFM Open License v1.0 (a copy is in this repository and in the pack). This pack is a derivative of LiquidAI/LFM3-1.7B under the same license; "LFM" and "Liquid AI" name the model's origin only. Answers come from an AI model and can be wrong.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Maximilian228/Geode-LFM3-1.7B

Finetuned
(89)
this model