【招聘】 MiniMax 视频模型全球开源生态
MiniMax正在寻找真正热爱开源、熟悉 AI 视频生成技术与开发者生态的 Open Source Ecosystem Lead，负责 MiniMax 视频模型的全球开源生态建设。
欢迎转发 https://vrfi1sk8a0.jobs.feishu.cn/index/position/7671161153523157311/detail
MiniMax正在寻找真正热爱开源、熟悉 AI 视频生成技术与开发者生态的 Open Source Ecosystem Lead，负责 MiniMax 视频模型的全球开源生态建设。
欢迎转发 https://vrfi1sk8a0.jobs.feishu.cn/index/position/7671161153523157311/detail
Can the license be changed to something like Apache 2.0 or OpenMDW 1.1 and can we get an undistilled base model, please!
已投递~
Hey, thanks for this model! It’s really good :)
Would it be possible to release the undistilled bf16 checkpoint to make LoRA training more effective? Then it would be truly open source 😃
On the undistilled checkpoint request — one measured data point that may support it.
On the pruned base (turbo LoRA), LoRA application is not fully effective, and the
failure is concentrated in exactly the path that step distillation rewrites. Of 259
LoRA-targeted modules, 208 applied and 51 were skipped — all of them adaln_proj.linear.
The cause is a dimension mismatch on that path: the pruned base exposes input dim 8,
the LoRA expects 2688. The LoRA metadata itself carries
incompatible_base: MiniMax-H3 pruned_*. Net effect I measured: 208/259 = 80.3% of
modules take the LoRA.
Precision matters here: that is an inference-side number on the pruned base, not a
training result, so treat it as corroboration rather than proof. But the skipped set
being AdaLN is consistent with your point that distillation is what limits LoRA
usefulness. Note also that this failure does not abort generation — it logs an ERROR
per step, which is easy to mistake for noise.
One flag for whoever ships an undistilled bf16: the memory floor matters more than the
VRAM number. On a 16GB-RAM / 8GB-VRAM box the working set I measured is ~34.5 GiB
(19.5 GiB main model + 14.96 GiB nvfp4 text encoder) against 15.26 GiB usable RAM, so it
pages. The tell is not utilization — that still reads 99% — it is board power
flat-lining at 34-36 W instead of the normal 64-98 W. Training-oriented builds will land
on machines like this first.
Happy to share the logs and config behind both numbers if useful.