Add 768x480 vision towers (Qwen3.5, Gemma 4); update 512x320 Qwen tower

New Core ML model packages:
- qwen3_5_mlx_4b_vision_vid_768x480_w8 Qwen3.5-4B vision tower @ 768x480
- gemma4_mlx_e4b_vision_768x480_w8 Gemma 4 (e4b) vision tower @ 768x480

Re-convert the existing Qwen3.5-4B vision tower @ 512x320 to the
tensor_video_rgb variant: inputs are now two planar RGB fp16 tensors
(1,3,320,512) in 0..255 with x/127.5-1 normalisation baked into the graph,
replacing the previous BGRA CVPixelBuffer image inputs. The new 768x480
Qwen build shares that I/O contract.

All new/updated towers are Apache-2.0 (Gemma 4 weights included -- unlike
earlier Gemma releases). Update README.md (I/O, shapes, precision,
conversion toolchain, SHA-256 checksums) and add a Gemma 4 upstream
component entry to THIRD_PARTY_LICENSES.md.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

README.md

@@ -21,7 +21,7 @@ VPIPE runtime can fetch a pinned, ready-to-run artifact instead of
 re-running the conversion.
 
 > **Licensing note:** the bundled models are **not** all under a single
-> license. Two are Apache-2.0 and one (Silero VAD) is MIT. See
+> license. Four are Apache-2.0 and one (Silero VAD) is MIT. See
 > [THIRD_PARTY_LICENSES.md](THIRD_PARTY_LICENSES.md) for the per-model
 > attribution and verbatim license texts.
 
@@ -45,26 +45,60 @@ written `WxH`.
 
 | Archive | Model | Task | License |
 | --- | --- | --- | --- |
-| `qwen3_5_mlx_4b_vision_vid_512x320_w8.tar` | Qwen3.5-4B vision tower (Core ML) | Vision encoder for a multimodal LLM | Apache-2.0 |
+| `qwen3_5_mlx_4b_vision_vid_512x320_w8.tar` | Qwen3.5-4B vision tower @ 512×320 (Core ML) | Vision encoder for a multimodal LLM | Apache-2.0 |
+| `qwen3_5_mlx_4b_vision_vid_768x480_w8.tar` | Qwen3.5-4B vision tower @ 768×480 (Core ML) | Vision encoder for a multimodal LLM | Apache-2.0 |
+| `gemma4_mlx_e4b_vision_768x480_w8.tar` | Gemma 4 (e4b) vision tower @ 768×480 (Core ML) | Vision encoder for a multimodal LLM | Apache-2.0 |
 | `yolox_l_1024x640_w8.tar` | YOLOX-L (Core ML) | Object detection | Apache-2.0 |
 | `silero-vad-unified-v6.tar` | Silero VAD v6, unified (Core ML) | Voice activity detection | MIT |
 
 ### `qwen3_5_mlx_4b_vision_vid_512x320_w8`
 
-Qwen3.5-4B vision tower (`variant=ane_v1_video`), the vision encoder that
-turns video frames into the embedding sequence consumed by the Qwen3.5-4B
-language model.
+Qwen3.5-4B vision tower (`variant=tensor_video_rgb`), the vision encoder
+that turns a pair of video frames into the embedding sequence consumed by
+the Qwen3.5-4B language model.
 
-- **Task:** image/video → vision feature embeddings (multimodal LLM front-end).
+- **Task:** video → vision feature embeddings (multimodal LLM front-end).
 - **Inputs:** `image0`, `image1` — the two consecutive video frames that
-  form one temporal patch, each **512×320 (W×H)**. Fed as a BGRA8888
-  `CVPixelBuffer`; the graph drops alpha and applies a `1/127.5` scale and
-  `-1` bias, yielding a BGR tensor in `[-1, 1]`.
-- **Output:** `image_features` — vision features of shape `(160, 2560)`.
+  form one temporal patch, each a planar RGB fp16 tensor `(1, 3, 320, 512)`
+  (`1×3×H×W`, i.e. 512×320 W×H) with raw pixels in `0..255`. The
+  `x/127.5 - 1` normalisation is baked into the graph (applied on the ANE).
+- **Output:** `image_features` — vision features of shape `(160, 2560)`, fp16.
 - **Precision:** 8-bit weight quantization, FP16 compute.
 - **Upstream:** [Qwen3.5](https://huggingface.co/Qwen) (Qwen team, Alibaba Cloud) — Apache-2.0.
 - **Conversion:** coremltools 9.0 from TorchScript (`torch==2.12.0`).
 
+### `qwen3_5_mlx_4b_vision_vid_768x480_w8`
+
+The same Qwen3.5-4B vision tower (`variant=tensor_video_rgb`) at a higher
+768×480 input — identical I/O contract to the 512×320 build, with a longer
+output token sequence.
+
+- **Task:** video → vision feature embeddings (multimodal LLM front-end).
+- **Inputs:** `image0`, `image1` — the two consecutive video frames that
+  form one temporal patch, each a planar RGB fp16 tensor `(1, 3, 480, 768)`
+  (`1×3×H×W`, i.e. 768×480 W×H) with raw pixels in `0..255`. The
+  `x/127.5 - 1` normalisation is baked into the graph (applied on the ANE).
+- **Output:** `image_features` — vision features of shape `(360, 2560)`, fp16.
+- **Precision:** 8-bit weight quantization, FP16 compute.
+- **Upstream:** [Qwen3.5](https://huggingface.co/Qwen) (Qwen team, Alibaba Cloud) — Apache-2.0.
+- **Conversion:** coremltools 9.0 from TorchScript (`torch==2.12.0`).
+
+### `gemma4_mlx_e4b_vision_768x480_w8`
+
+Gemma 4 (e4b) vision tower, converted to Core ML. It pools and projects the
+visual features into the language model's 2560-dim text embedding space,
+emitting "soft tokens" ready to splice into the Gemma 4 prompt.
+
+- **Task:** image → LLM soft tokens (multimodal LLM front-end).
+- **Input:** `image` — a single planar RGB fp16 MLMultiArray
+  `[1, 3, 480, 768]` (`1×3×H×W`, i.e. 768×480 W×H), raw pixels in `0..255`
+  (channel order R, G, B). Preprocessing is baked in: `x/255` then
+  `2*(x - 0.5)`.
+- **Output:** `soft_tokens` — shape `[1, 160, 2560]`, fp16.
+- **Precision:** int8 weight-only quantization (per-block), FP16 compute.
+- **Upstream:** [Gemma 4 (e4b)](https://huggingface.co/google/gemma-4-E4B) (Google DeepMind) — Apache-2.0.
+- **Conversion:** coremltools 9.0 from TorchScript (`torch==2.12.1`).
+
 ### `yolox_l_1024x640_w8`
 
 YOLOX-L anchor-free object detector, converted to Core ML at a fixed
@@ -100,7 +134,9 @@ Core ML conversion by FluidInference.
 SHA-256 checksums of the archives:
 
 ```
-3940398c2518ac5f01c09f5f5eab868c2b217ab53c7d9a145d2387e76251b5d5  qwen3_5_mlx_4b_vision_vid_512x320_w8.tar
+ed27924a97a873a019d1c8a1e1f9826bd30da2588683843d6c59cfb0b2289f30  qwen3_5_mlx_4b_vision_vid_512x320_w8.tar
+3e928ea0c3e6848896278b68436d46f1aa306d95d8ca4a1faadfad354d705b87  qwen3_5_mlx_4b_vision_vid_768x480_w8.tar
+17bebe9fdb416a760df4d00ed39c7edf296ba768d54139c8183c1323da646d3e  gemma4_mlx_e4b_vision_768x480_w8.tar
 7f7bf85947bb228f978470284e16a9d00c906a6cb07359f9a51c68806fa9c609  yolox_l_1024x640_w8.tar
 ee14d006d1cea87f15a4f0ac5f3e1e00b80a10c40eca855e8ff6741f389b4423  silero-vad-unified-v6.tar
 ```

THIRD_PARTY_LICENSES.md

@@ -6,13 +6,16 @@ component and the archive it ships in, identifies the upstream source,
 reproduces the upstream copyright statement, and points to the verbatim
 license text required by that component's license.
 
-Two of the three models are licensed under the Apache License, Version 2.0;
-the third (Silero VAD) is licensed under the MIT License. The verbatim
-license texts are reproduced in the appendices at the end of this file.
+Every model except one is licensed under the Apache License, Version 2.0;
+the remaining one (Silero VAD) is licensed under the MIT License. The
+verbatim license texts are reproduced in the appendices at the end of this
+file.
 
 | Archive | Component | License |
 | --- | --- | --- |
 | `qwen3_5_mlx_4b_vision_vid_512x320_w8.tar` | Qwen3.5 vision tower | Apache-2.0 ([Appendix A](#appendix-a-apache-license-version-20)) |
+| `qwen3_5_mlx_4b_vision_vid_768x480_w8.tar` | Qwen3.5 vision tower | Apache-2.0 ([Appendix A](#appendix-a-apache-license-version-20)) |
+| `gemma4_mlx_e4b_vision_768x480_w8.tar` | Gemma 4 vision tower | Apache-2.0 ([Appendix A](#appendix-a-apache-license-version-20)) |
 | `yolox_l_1024x640_w8.tar` | YOLOX | Apache-2.0 ([Appendix A](#appendix-a-apache-license-version-20)) |
 | `silero-vad-unified-v6.tar` | Silero VAD (FluidInference Core ML conversion) | MIT ([Appendix B](#appendix-b-mit-license)) |
 
@@ -20,10 +23,11 @@ license texts are reproduced in the appendices at the end of this file.
 
 ## Qwen3.5 (vision tower)
 
-The archive `qwen3_5_mlx_4b_vision_vid_512x320_w8.tar` contains a Core ML
-conversion of the **Qwen3.5-4B** vision tower (the multimodal vision
-encoder), derived from the Qwen3.5 model series published by the Qwen team
-at Alibaba Cloud (<https://huggingface.co/Qwen>). The weights are
+The archives `qwen3_5_mlx_4b_vision_vid_512x320_w8.tar` and
+`qwen3_5_mlx_4b_vision_vid_768x480_w8.tar` contain Core ML conversions of
+the **Qwen3.5-4B** vision tower (the multimodal vision encoder) at two input
+resolutions, derived from the Qwen3.5 model series published by the Qwen
+team at Alibaba Cloud (<https://huggingface.co/Qwen>). The weights are
 8-bit-quantized and the graph is exported to Core ML; the model is otherwise
 a derivative of the upstream Qwen3.5 weights.
 
@@ -46,6 +50,37 @@ upstream at
 
 ---
 
+## Gemma 4 (vision tower)
+
+The archive `gemma4_mlx_e4b_vision_768x480_w8.tar` contains a Core ML
+conversion of the **Gemma 4 (e4b)** vision tower (the multimodal vision
+encoder that projects image features into the language model's text
+embedding space), derived from the Gemma 4 model series published by Google
+DeepMind (<https://huggingface.co/google/gemma-4-E4B>). The weights are
+int8 weight-only quantized and the graph is exported to Core ML; the model
+is otherwise a derivative of the upstream Gemma 4 weights.
+
+Gemma 4 open-weight models are distributed under the Apache License,
+Version 2.0. (This differs from earlier Gemma releases, which used the
+custom Gemma Terms of Use.)
+
+### Copyright notice
+
+> Copyright Google LLC
+
+The upstream model card declares the license as Apache-2.0; Google does not
+publish a per-file copyright line for the Gemma 4 weights.
+
+### License
+
+Apache License, Version 2.0. The verbatim text is reproduced in
+[Appendix A](#appendix-a-apache-license-version-20), and is available
+upstream at
+<https://ai.google.dev/gemma/docs/gemma_4_license> and
+<https://www.apache.org/licenses/LICENSE-2.0>.
+
+---
+
 ## YOLOX
 
 The archive `yolox_l_1024x640_w8.tar` contains a Core ML conversion of

gemma4_mlx_e4b_vision_768x480_w8.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:17bebe9fdb416a760df4d00ed39c7edf296ba768d54139c8183c1323da646d3e
+size 165332992

qwen3_5_mlx_4b_vision_vid_512x320_w8.tar

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3940398c2518ac5f01c09f5f5eab868c2b217ab53c7d9a145d2387e76251b5d5
-size 333077504
+oid sha256:ed27924a97a873a019d1c8a1e1f9826bd30da2588683843d6c59cfb0b2289f30
+size 333078016

qwen3_5_mlx_4b_vision_vid_768x480_w8.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3e928ea0c3e6848896278b68436d46f1aa306d95d8ca4a1faadfad354d705b87
+size 334001152