immanuelpeter

Ship the parent-order forward (modeling_pixtral_large4.py) and correct the card; weights unchanged (vision-tower-bench 5139e1a)

3cd67e8 verified 7 days ago

Raw History Blame Contribute Delete

1.24 kB

Attribution notice: 2026 - Mistral

Weights: mistralai/Mistral-Large-4-1T-A52B at f12dabdc0c1e99e7f7924461a6f308859b05af4d,
shard consolidated-00053-of-00054.safetensors
(SHA256 d3289cad598653fec78adfd49f939504f92c4c0e9d8b9e66c8750b2e917dde90).
LICENSE is the parent's Mistral Ecosystem Open-Weight License, copied byte for byte.
It is a modified MIT license. Its Section 2 requires a separate commercial license from
Mistral before a Model as a Service business uses the Model or derivative works for any
commercial purpose.

Only the 272 vision encoder and 4 multimodal projector tensors are included. No language
model weights. Tensor values are unchanged BF16. Names follow HF Pixtral and Mistral3, and
the rows of each attention wq and wk are reordered per head from Mistral's interleaved
rotary layout to HF's rotate-half layout. export_manifest.json records the SHA256 of
every source tensor before renaming.

modeling_pixtral_large4.py, projector.py and examples/inference.py come from the
vision-tower-bench repository under the MIT license in LICENSE-CODE. The modeling
file subclasses transformers' PixtralVisionModel (Apache-2.0) and follows the
numeric order of mistral-inference (Apache-2.0); it copies no code from either.

	Attribution notice: 2026 - Mistral

	Weights: mistralai/Mistral-Large-4-1T-A52B at f12dabdc0c1e99e7f7924461a6f308859b05af4d,
	shard consolidated-00053-of-00054.safetensors
	(SHA256 d3289cad598653fec78adfd49f939504f92c4c0e9d8b9e66c8750b2e917dde90).
	LICENSE is the parent's Mistral Ecosystem Open-Weight License, copied byte for byte.
	It is a modified MIT license. Its Section 2 requires a separate commercial license from
	Mistral before a Model as a Service business uses the Model or derivative works for any
	commercial purpose.

	Only the 272 vision encoder and 4 multimodal projector tensors are included. No language
	model weights. Tensor values are unchanged BF16. Names follow HF Pixtral and Mistral3, and
	the rows of each attention wq and wk are reordered per head from Mistral's interleaved
	rotary layout to HF's rotate-half layout. export_manifest.json records the SHA256 of
	every source tensor before renaming.

	modeling_pixtral_large4.py, projector.py and examples/inference.py come from the
	vision-tower-bench repository under the MIT license in LICENSE-CODE. The modeling
	file subclasses transformers' PixtralVisionModel (Apache-2.0) and follows the
	numeric order of mistral-inference (Apache-2.0); it copies no code from either.