mlboydaisuke
/

FLUX.2-klein-4B-CoreAI

@@ -24,8 +24,8 @@ This model has no row on [DeviceMark](https://devicemark.github.io/), the on-dev
 > **2026-10-10 re-export, two folders.** `macos-fp16/` and `macos-int8/` were exported with apple/coreai-models
 > 97a14be and load with its diffusion pipeline. The files at the repo root are the 2026-07-20 export: current
-> apple/coreai-models stops on them with "unsupported metadata_version". They stay for the zoo's CoreAIImageGen app,
-> which pins a fork of that time.
 # FLUX.2 klein 4B — Core AI
@@ -45,7 +45,7 @@ iPhone 17 Pro, over the ~6.1 GB a 12 GB iPhone gives one app.
 | --- | --- | --- | --- | --- |
 | `macos-fp16/` | fp16, one asset per graph | text-to-image | 19 | 14,159,882,995 |
 | `macos-int8/` | int8 per-block 32 on the text encoder and the transformer, fp16 VAEs | text-to-image, image-to-image | 25 | 7,777,120,667 |
-| repo root (2026-07-20) | int4 per-block 32, plus the in-context edit transformers | the zoo's CoreAIImageGen app only | 34 | 10,854,216,650 |
 Text-to-image reads 16 files of a folder: 14,090,986,305 bytes from `macos-fp16/`, 7,540,032,209 bytes from
 `macos-int8/`. The VAE encoders and the half-size VAEs are only for image-to-image and tiled decoding.
@@ -164,9 +164,9 @@ re-renders the whole frame.
 It is the same DiT graph exported at a longer sequence: the output latent (time index `T=0`) concatenated with the
 reference image's latent tokens (`T=10`), so the transformer attends to the reference while denoising the output.
-`Transformer_edit_2ref.aimodel` takes two references (`T=10`, `T=20`). Running them needs a runtime that drives this
-path: the zoo's [**CoreAIImageGen**](https://github.com/john-rocky/coreai-model-zoo/tree/main/apps/CoreAIImageGen) app
-pins the `john-rocky/coreai-models` fork with `Flux2Pipeline.editImages` and exposes it as the **Edit** tab. These files
 are int4 and use the 2026-07-20 layout, so apple/coreai-models 97a14be does not load them.
 int4, ~25 s for a 1024 edit and ~43 s for a two-reference edit on a Mac GPU (4 steps, guidance 1.0; numbers from

 > **2026-10-10 re-export, two folders.** `macos-fp16/` and `macos-int8/` were exported with apple/coreai-models
 > 97a14be and load with its diffusion pipeline. The files at the repo root are the 2026-07-20 export: current
+> apple/coreai-models stops on them with "unsupported metadata_version". No app reads them since 2026-10-10: the zoo's
+> CoreAIImageGen app moved to the two folders. They stay for the record.
 # FLUX.2 klein 4B — Core AI
 | --- | --- | --- | --- | --- |
 | `macos-fp16/` | fp16, one asset per graph | text-to-image | 19 | 14,159,882,995 |
 | `macos-int8/` | int8 per-block 32 on the text encoder and the transformer, fp16 VAEs | text-to-image, image-to-image | 25 | 7,777,120,667 |
+| repo root (2026-07-20) | int4 per-block 32, plus the in-context edit transformers | no app since 2026-10-10 (kept for the record) | 34 | 10,854,216,650 |
 Text-to-image reads 16 files of a folder: 14,090,986,305 bytes from `macos-fp16/`, 7,540,032,209 bytes from
 `macos-int8/`. The VAE encoders and the half-size VAEs are only for image-to-image and tiled decoding.
 It is the same DiT graph exported at a longer sequence: the output latent (time index `T=0`) concatenated with the
 reference image's latent tokens (`T=10`), so the transformer attends to the reference while denoising the output.
+`Transformer_edit_2ref.aimodel` takes two references (`T=10`, `T=20`). Running them needs the runtime that drove this
+path, the `john-rocky/coreai-models` branch `flux2-in-context-edit` (bc48d3d); the zoo's [**CoreAIImageGen**](https://github.com/john-rocky/coreai-model-zoo/tree/main/apps/CoreAIImageGen)
+app used it until 2026-10-10 and now runs the stock one-reference image-to-image as its **Edit** tab. These files
 are int4 and use the 2026-07-20 layout, so apple/coreai-models 97a14be does not load them.
 int4, ~25 s for a 1024 edit and ~43 s for a two-reference edit on a Mac GPU (4 steps, guidance 1.0; numbers from