| # Launch announcement draft |
|
|
| MiniMax H3 Image on Apple silicon, with native MLX/Metal inference. |
|
|
| Four weight choices: original BF16, lossless packed BF16, Q8 and Q4. Download |
| one variant and reuse the shared components. Includes six-step generation, |
| reference-image editing, copyable quick starts, original image comparisons, |
| numerical audits and benchmark records. |
|
|
| On the tested M5 Pro / 64 GiB, the recorded exact tuned profiles measured |
| about 62–64 seconds for warm 2048-square generation. Details, measurement |
| boundaries and fidelity tradeoffs are documented. The native text encoder is |
| a separate large download and can be reused from an existing installation. |
|
|
| Independent community port by Devin Lai. Code is Apache-2.0; weights retain |
| MiniMax's community licenses and access restrictions. No hosted service. |
|
|
| Start here: https://huggingface.co/devin-lai/MiniMax-H3-Image-mlx |
|
|
| Share a result with the variant, prompt, seed, image size, chip/memory and |
| generated JSON record. Reports from more Macs will help make support better. |
|
|
| 简体中文：MiniMax H3 Image 的 Apple 芯片 MLX 本地推理社区移植。 |
| 按需选择 BF16、无损 packed12、Q8 或 Q4，复用公共组件。 |
| 包含图像编辑、原始对比图和可复现实测。模型保留 MiniMax 社区许可及相关限制。 |
|
|
