| 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. |
|
|
