Hugging Face Releases 3 New AI Models: Better Speech, Vision, and Speed
New AI models are faster and smarter — here's what they mean for you.
Hugging Face Transformers released v5.16.0, adding several new models: Qwen4-Exp, GraniteSpeech5, Step-3.7-Flash, and CohereCompass. Qwen4-Exp builds on Qwen3.5's hybrid text and multimodal architecture, combining a residual design, sparse attention, and per-layer embeddings to improve inference efficiency for long-context workloads. GraniteSpeech5 is a lightweight (~470M parameters) conformer encoder for speech recognition, using CTC and a single forward pass. Step-3.7-Flash is a 198B-parameter sparse mixture-of-experts vision-language model with a 1.8B vision encoder, capable of native image understanding. The release notes also list CohereCompass, but that entry is cut off in the source.
- Hugging Face added three new AI models: one for long documents, one for speech recognition, and one for understanding images.
- The new speech model is small enough to run on phones and everyday devices, potentially making voice assistants faster and more private.
- The vision-language model can describe images, and it uses energy efficiently — but its creators haven't shared full technical details yet.
Why It Matters
These updates mean AI apps you use will get faster, cheaper, and better at understanding speech and images.