refactor(xiaomi): all mimo models extend from the xiaomi root provider#1631
refactor(xiaomi): all mimo models extend from the xiaomi root provider#1631juls0730 wants to merge 1 commit intoanomalyco:devfrom
Conversation
juls0730
force-pushed
the
refactor/mimo-extends
branch
9 times, most recently
from
17faba2 to
a5ee8bd
Compare
|
I have also updated the open weights declarations for mimo v2.5 and pro https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro https://huggingface.co/XiaomiMiMo/MiMo-V2.5 however V2 pro and omni remain closed |
juls0730
force-pushed
the
refactor/mimo-extends
branch
5 times, most recently
from
1f3c575 to
fd55568
Compare
|
I corrected the input modalities on the mimo V2.5 models, in error I originally had pro and non-pro flip flopped, oops. |
|
Hey @rekram1-node id love if you could take a look at this pr 🙏 |
|
I'd prefer u do this one by one ideally per provider I think there are multiple regressions in this pr |
| tool_call = true | ||
| knowledge = "2025-01-01" | ||
| open_weights = false | ||
|
|
There was a problem hiding this comment.
regression in pricing right
There was a problem hiding this comment.
I think no regression. All that changed was that cache_read was added, which according to the zenmux website is correct, unless you mean something else?
| [modalities] | ||
| input = ["text", "image", "video"] | ||
| output = ["text"] | ||
| [extends] |
There was a problem hiding this comment.
regression in several items
There was a problem hiding this comment.
modalities
V2 Pro does not support image or video input.
cost
Same as with V2 omni, pricing is correct all that changes is its more detailed with context over 200k and has cache_read cost
limit
it seems possible that the limit is actually a regression, the zenmux page lists it as 256k under the xiaomi provider, which seems implausible to me, but I suppose a valid regression. Omni also suffers from this
juls0730
force-pushed
the
refactor/mimo-extends
branch
from
fd55568 to
dfaba58
Compare
juls0730
force-pushed
the
refactor/mimo-extends
branch
from
dfaba58 to
048b94b
Compare
2 participants
Refactor all MiMo models to extend from the Xiaomi provider as a single source of truth. Draft PR until I can 100% verify I have not made any errors. This PR is the first of potentially a series of commits intended to create single sources of truths for most models behind wrappers.
🤖 PR generated with the help of MiMo V2.5 Pro