File size: 1,303 Bytes
3da88e3 | 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 | # 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。
|