Transformers v5.15.0 adds model support and changes kernel defaults
Transformers v5.15.0 lists support for Muse Glimmer, GraniteSWA, GraniteMoeSWA, A.X-K1, A.X-K2, and Cosmos3 Edge. Its principal upgrade concern is that kernels for linear-attention models are now opt-in, requiring explicit configuration for users who previously depended on automatic selection.
Hugging Face Transformers v5.15.0 adds support for several model families and introduces a compatibility change for linear-attention kernels. The release record says that kernels are no longer enabled automatically for affected models, making upgrade testing important for projects that relied on prior automatic selection.
01
What we know now
- 01
[1] Hugging Face Transformers, official release record, “Release: v5.15.0”, published 10 August 2026: https://github.com/huggingface/transformers/releases/tag/v5.15.0.
- 02
The supplied evidence is a first-party release record and does not include an independent recheck.
02
Why this matters for Armenia
Developers in Armenia using the open-source Transformers library can use this release summary to assess new model options and prepare for the changed kernel configuration before upgrading.
03
Newly listed support: GraniteSWA and GraniteMoeSWA
The release record lists support for both GraniteSWA variants.Newly listed support: A.X-K1 and A.X-K2
The release record lists both A.X-K model variants as additions.Newly listed support: Cosmos3 Edge
Cosmos3 Edge is included among the newly supported models.Kernel behavior for listed linear-attention models
Kernel use is described as opt-in, so automatic selection should not be assumed after upgrading.Status labels summarize statements in the official v5.15.0 release record.
04
Additional model families enter the library
The v5.15.0 release record identifies five newly supported model lines or variants: GraniteSWA, GraniteMoeSWA, A.X-K1, A.X-K2, and Cosmos3 Edge. For developers, the immediate implication is that these architectures can be evaluated through Transformers support documented with this release. [1]
- GraniteSWA and GraniteMoeSWA are listed as newly supported.
- A.X-K1, A.X-K2, and Cosmos3 Edge are also listed as new model additions.
- The release record presents these as library support additions; it does not, in the supplied material, provide performance comparisons or implementation guidance for each family.
05
Muse Glimmer support includes a multimodal 30B model
According to the release record, Meta Muse Glimmer is a 30B-parameter multimodal model composed of a 2B ViT-style vision encoder and a 28B text decoder. The same record states that it is available under the Apache 2.0 licence. These details are first-party release claims rather than independently verified assessments. [1]
- The release record describes Muse Glimmer as multimodal.
- It specifies a 2B ViT-style vision encoder and a 28B text decoder, for 30B parameters in total.
- It says the model is released under the Apache 2.0 licence.
06
Kernel selection becomes an explicit compatibility task
The release record flags a breaking change: kernels are now opt-in for linear-attention models. It specifically identifies Mamba, GDN, and convolution-only models as examples, and says users who relied on automatic kernel selection must explicitly enable kernels. Teams should therefore treat kernel configuration as an upgrade requirement, while consulting the full release documentation for exact settings because the supplied excerpt is incomplete. [1]
- Kernels are described as opt-in rather than mandatory for linear-attention models.
- The release explicitly names Mamba, GDN, and convolution-only models as examples.
- Projects that relied on automatic kernel selection now need explicit enablement.
- The supplied material does not state the full configuration procedure or all affected migration cases.
07
Upgrade checks
Before moving to v5.15.0, developers can make the review focused and reversible.
- 01
Check whether any project uses Mamba, GDN, or convolution-only linear-attention models.
- 02
Review code and deployment settings that previously depended on automatic kernel selection.
- 03
Explicitly enable the required kernels where applicable, then test representative inference or training workloads.
- 04
Evaluate the newly supported model families against application requirements and licensing needs.
08
Limits of this edition
All claims are drawn from Hugging Face’s first-party release record; no independent verification was supplied.
The supplied release material does not provide the full migration guidance for the kernel change.
The release record’s model descriptions and licensing information are attributed claims and should be checked against the linked release documentation before deployment.
SRC
Source desk
Direct links to the material behind this selection. Seeing the source matters as much as reading the synthesis.
Keep exploring
20 Aug 2026
Hugging Face record identifies DeepSeek-V4-Pro-0813 as a text-generation repository
The supplied Hugging Face registry record establishes that deepseek-ai/DeepSeek-V4-Pro-0813 is listed under deepseek-ai, with text-generation, transformers, and mit metadata labels. It records creation and modification timestamps on 13 August 2026. Capability, access, file-content, and detailed license questions remain unresolved. [1]
17 Aug 2026
Run a local language model on Apple Silicon, without a cloud account
A safe first MLX LM run, with realistic model sizing, copyable commands, and checks for privacy, downloads, and model terms.
18 Aug 2026
OpenAI Python SDK v2.49.0 sets a Python 3.10 requirement
OpenAI Python SDK v2.49.0 declares Python 3.10 as its required Python version and lists automated version reviews. The supplied release record does not explain migration steps or the exact behavior on older Python releases.