File size: 2,517 Bytes
a46a994 e20b23a a46a994 e20b23a a46a994 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 | Independent Core AI conversion of LiquidAI/d1-3B for text and image decisions.
Upstream: https://huggingface.co/LiquidAI/d1-3B
Revision: 051bcc464b01b9f92942b364d9586b0ef5912432
The upstream checkpoint, upstream code and converted weights are governed by the
LFM Open License v1.0. A copy of the license accompanies this distribution.
The conversion uses 1x1 convolution projections, activation scaling,
tanh-form SiLU and a split FFN. The language backbone and SigLIP2 vision tower
are exported as FP16 Core AI assets for the Apple Neural Engine. Tokenization,
image preprocessing, token-embedding lookup, the FP32 multimodal projector,
and option readout run on the host CPU. The model has not been retrained.
This project is not affiliated with or endorsed by Liquid AI.
Independent Core AI conversions of vllm-sr/d3-nano and vllm-sr/d3-mini,
fine-tuned from Qwen/Qwen3.5-2B and Qwen/Qwen3.5-4B, for text and image decisions.
Upstream: https://huggingface.co/vllm-sr/d3-nano, https://huggingface.co/vllm-sr/d3-mini
Revisions are pinned in src/tinkerbox_ane/d3/__init__.py and recorded in exported manifests.
Their checkpoints, upstream code and converted weights are governed by the
Apache License 2.0 (LICENSES/Apache-2.0.txt). The language backbone runs as
FP16 Core AI stages and the vision tower as an FP16 Core AI asset on the Apple
Neural Engine; tokenization, image preprocessing, position tables,
token-embedding lookup and option readout run on the host CPU. The models
have not been retrained. This project is not affiliated with or endorsed by
vllm-sr or Qwen.
OmniVoice integration:
Upstream code: https://github.com/k2-fsa/OmniVoice (Apache-2.0).
Checkpoint: https://huggingface.co/k2-fsa/OmniVoice
Revision: c5fdb5ccb189668d56333f77ba2629f4cd7535f4
OmniVoice checkpoint and derived weights are licensed CC-BY-NC.
The Qwen3 transformer is converted independently; CPU speech processing and
sampling use the separately installed upstream omnivoice package.
VoxCPM2 speech integration:
Upstream: https://github.com/OpenBMB/VoxCPM
Source revision: cce58a5b59303c9bd63d12f23afe6a49a4a80c59
Checkpoint: https://huggingface.co/openbmb/VoxCPM2
Revision: 32279effe8c19989596f05d353d1447f51d9e915
Copyright 2026 OpenBMB. Code and weights are Apache-2.0 licensed;
see LICENSES/VoxCPM-Apache-2.0.txt. The AudioVAE V2 codec and causal streaming decoder
are retained from upstream. MiniCPM transformers are independently converted
for ANE; task conditioning and sampling follow the pinned upstream contract.
|