LFM3 model packages

This repository contains LFM3 model packages for ZETIC's device-specific runtimes. The Qualcomm packages are for the ZeticMLange Hexagon SDK 0.1.0. The Apple packages are for the separate LFM3 iOS SDK, currently under development. The SDKs and model packages are distributed separately.

Layout

qualcomm/<model>/<target>/       Qualcomm Hexagon NPU packages
apple/lfm3-1.7b/ios/             LFM3-1.7B iOS model bundle
apple/lfm3-audio-2b/ios/          LFM3-Audio-2B iOS model bundle
apple/lfm3-vl-2b/ios/             LFM3-VL-2B iOS model bundle
apple/lfm3-omni-2b/ios/          LFM3-Omni-2B iOS model bundle

The package formats and SDKs are different; use the package for your target platform.

Qualcomm Hexagon

These packages run the LFM3 family on the Qualcomm Hexagon NPU through Ixian Hexagon, ZETIC's own Hexagon runtime. The SDK itself (library, headers and CLIs) is shared separately as a tarball. Each package folder contains manifest.json, SHA256SUMS, README.md, the tokenizer and the weights.

Which target to use

Device	SoC	Target	SDK tarball
Galaxy S25	SM8750	v79	zetic-mlange-hexagon-0.1.0-android-arm64.tar.gz
Galaxy S26 Ultra	SM8850	v81	zetic-mlange-hexagon-0.1.0-android-arm64.tar.gz
Dragonwing IQ9 (2 NSPs)	QCS9075	v73	zetic-mlange-hexagon-0.1.0-linux-arm64-qcom.tar.gz

Models

• lfm3-1.7b: text
• lfm3-vl-2b: text, image, video
• lfm3-audio-2b: text, audio (ASR), speech output
• lfm3-omni-2b: text, image, audio, video, speech output
• d1-2b and d1-omni-2b: decisions, through the Decision API

A Qualcomm package is 1.2–3.6 GB.

Download

Download only the models and targets you need. For example, lfm3-1.7b for the Galaxy S26 Ultra:

hf download ZETIC-ai/LFM3 --include "qualcomm/lfm3-1.7b/v81/*" --local-dir .
(cd qualcomm/lfm3-1.7b/v81 && sha256sum -c SHA256SUMS)

Then pass the model directory, the one that holds the target folders, to the SDK:

LLM llm = LLM::fromPath(".../qualcomm/lfm3-1.7b");   // picks v79 / v81 / v73 for the device

The Hexagon SDK README.md has the rest.

Apple iOS

LFM3-Omni-2B for iOS is an approximately 4.1 GB directory bundle containing Core ML and Metal assets for text, audio and image input and speech output. It uses the separate LFM3 iOS SDK (PR #84), which is under development; it is not a Hexagon SDK package.

The other iOS bundles are LFM3-1.7B, LFM3-Audio-2B, and LFM3-VL-2B. Each bundle uses its corresponding SDK session and has a separate package ID and license alongside it.

The repository is gated. Authenticate with a Hugging Face account that has access, then download the bundle and its license:

hf download ZETIC-ai/LFM3 \
  --include 'apple/lfm3-omni-2b/ios/**' \
  --include 'apple/lfm3-omni-2b/LICENSE' \
  --local-dir .

Copy apple/lfm3-omni-2b/ios to app-accessible storage, retaining all files and relative paths. This directory is the model bundle; do not add files inside it. Open it from the SDK's native C++ / Objective-C++ API using its trusted package ID:

#include "lfm3_omni_session.h"
#include <string>

std::string error;
lfm3_ios::OmniSession session;
if (!session.open(bundle_directory,
                  "43669fe025f8ccdf5aa5a011f3ad4e610c922cc28f79d790f874e737f7661dd6",
                  error)) {
    // Handle error.
}

The package ID verifies bundle integrity; it is not a publisher signature. Keep the bundle unchanged while the session is open. This export uses a fixed 1,536-row prefill; see the iOS SDK documentation and example in PR #84 for input limits, building and inference. The SDK example has an iPhone text smoke test; its media examples have been checked by compiling and linking, not by that device smoke test. For the other iOS model bundles, consult their linked READMEs for fixed prefill sizes, model sources and package IDs.

License

The LFM3 weights are under the LFM Open License v1.0 (Liquid AI). Each Apple model directory includes its original LICENSE and a README identifying the source model and conversion.