The Wire · Showcase
TRANSFORMERS 5.14 SHIPS WITH INKLING MULTIMODAL MODEL
By RepoJournal · Filed · About Hugging Face
11 people shipped this
ArthurZucker
@ArthurZucker
1 cited
Pablo Montalvo
@molbap
1 cited
itazap
@itazap
1 cited
Mishig
@mishig25
1 cited
avishaigo-commits
@avishaigo-commits
1 cited
Vineeth Sai
@vineethsaivs
1 cited
tomaarsen
@tomaarsen
1 cited
sayakpaul
@sayakpaul
1 cited
tarekziade
@tarekziade
evalstate
@evalstate
delmalih
@delmalih
Hugging Face shipped transformers v5.14.0 with native support for Inkling, a 975B parameter multimodal model that processes text, image, and audio in a single forward pass.
The Inkling addition [1] [2] lands with fresh processor implementation and tokenizer fixes that address long-standing gaps in model coverage [3]. Parallel work hardened the tokenizer doc build pipeline [4], cutting build time through streamlined dependency management. On the infrastructure side, the hf-mcp-server received critical authentication hardening [5] that removes dangerous fallback token behavior for HTTP transports and enforces Host/Origin allowlisting by default. Sentence-transformers fixed a metric prefixing bug in ReciprocalRankFusionEvaluator [6] that broke trainer evaluation loops, and patched a torch 2.13 crash affecting bf16 computations on Windows [7]. Diffusers aligned snapshot_download behavior [8] to match online validation patterns when working offline, ensuring interrupted downloads don't silently accept corrupted caches.
Action items
- → Update to transformers 5.14.0 to enable Inkling multimodal support in production pipelines huggingface/transformers [plan]
- → Review hf-mcp-server HTTP authentication changes if running v0.3.31; upgrade to v0.3.32+ for production deployments huggingface/hf-mcp-server [immediate]
- → Upgrade sentence-transformers if using ReciprocalRankFusionEvaluator; metric access patterns were broken huggingface/sentence-transformers [immediate]
- → If running diffusers offline downloads, pull the snapshot_download fix to validate cached state huggingface/diffusers [plan]
References
- [1] Release v5.14.0 ↗ huggingface/transformers
- [2] Add Inkling model (#47347) huggingface/transformers
- [3] Fix models for which we don't have a dedicated tokenizer class, and the listed one is incorrect ↗ huggingface/transformers
- [4] Make doc builds faster (#2158) huggingface/tokenizers
- [5] Harden HTTP MCP request authentication ↗ huggingface/hf-mcp-server
- [6] Fix ReciprocalRankFusionEvaluator primary_metric not being prefixed (#3861) huggingface/sentence-transformers
- [7] [`tests`] Skip bf16 + Windows + CPU forwards, as they can WindowsError on torch 2.13 ↗ huggingface/sentence-transformers
- [8] align snapshot_download to respect hfh latest version ↗ huggingface/diffusers
