d1-Omni-2B for Core AI
Core AI conversion of LiquidAI/d1-Omni-2B,
checkpoint 4cccbfd5ba3405ee0f973134bd03e0d9d60ef1d8.
The model answers typed questions about text, documents, images, video and audio in one forward pass: choices, scores, boolean probabilities and image boxes. It does not generate chat or speech.
Compatibility
Requires iOS 27 or later and Core AI. h18p/ and h19p/ contain architecture-specific
GPU AOT bundles; use the set matching your device.
Confirm the architecture for your device with your Core AI toolchain. Architecture selection does not establish parity on every device.
Decoder context is 4,096 tokens, processed in chunks of 256. Vision uses chunks of 512 patches and mixing buckets from 512 to 8,192 in steps of 512. Audio uses chunks of 128 tokens and downsampling buckets of 32, 64, 128, 256, 512 and 1,024 tokens.
Files required by the app
| Path | Purpose |
|---|---|
<arch>/graphs/seg{0..6}.aimodelc |
Seven decoder segments holding the decoder weights |
<arch>/graphs/mix.aimodelc |
Full attention over the decoder's supported lengths |
<arch>/graphs/vseg{0..14}.aimodelc |
Fifteen vision segments |
<arch>/graphs/vconv_<P>.aimodelc |
Shared vision convolution mix, 16 patch lengths |
<arch>/graphs/vattn_<P>.aimodelc |
Shared vision attention mix, 16 patch lengths |
<arch>/graphs/audio_seg.aimodelc, down.aimodelc |
Audio encoder and downsampler |
<arch>/graphs/chunked.json |
Decoder shape and carry configuration |
<arch>/graphs/chunked_vision.json, vision_taps.f32 |
Shared vision configuration and convolution kernels |
<arch>/graphs/chunked_audio.json, meta.json |
Audio configuration and graph registry |
assets/tokenizer.json, calibration.json, embed.bf16 |
Tokenizer, calibrated readout and host embeddings |
<arch>/manifest.json |
Required runtime files, byte sizes and SHA-256 hashes |
There are 57 portable .aimodel bundles under aimodel/graphs/, and 57 compiled
.aimodelc bundles per architecture. The portable tree also includes the five graph
support files. assets/ is shared. The app downloads one architecture's graphs and the
host assets into Application Support. It verifies Hugging Face checksums, resumes interrupted
transfers, and places each completed directory atomically. Inference stays on the device.
The root manifest.json inventories the upload tree. provenance.json records source and
stripped hashes and structural validation. Original exports, compiler logs and private
strip records are excluded. All upload files, including binaries, are scanned for personal
paths. Debug locations are removed using conversion/d1_omni/strip_ship.py; operation counts
and coreai-build inspect summaries must match before and after. AOT binaries are compiled
from those stripped programs. This structural check is not a new device parity measurement.
Attribution and license
Original weights and assets: Liquid AI, Inc., under the included LFM Open License v1.0
(LICENSE). These files have been converted and modified; see NOTICE and provenance.json.
This is an independent port, not an official Liquid AI or Apple release.