Hugging Face Transformers v5 sheds TensorFlow, Flax, and 'slow' tokenizers in push for simplicity
Curated by the Inblix editorial team
The Hugging Face team launched Transformers v5 today, marking a deliberate shift from the sprawl of 400+ model architectures toward a simpler, more maintainable core. The library now sees over 3 million daily pip installs and has crossed 1.2 billion total installs — a staggering leap from the 20,000 daily installs recorded at v4’s launch. Community contributions have exploded too, with over 750,000 compatible model checkpoints now on the Hub, up from roughly 1,000 just a few years ago.
That scale came with real growing pains. The team’s answer is an opinionated pruning of legacy backends. Flax and TensorFlow support are being officially sunsetted, making PyTorch the sole framework. The old “Fast” and “Slow” tokenizer distinction is gone too — everything now runs on the tokenizers backend, with Sentencepiece and MistralCommon left as non-default alternatives. Image processors get the same treatment, surviving only in their “fast” torchvision-backed form.
Under the hood, the team is betting big on modularity to keep the architecture zoo manageable. They’re touting a new AttentionInterface that centralizes methods like FlashAttention and SDPA, pushing boilerplate out of individual model files. They’re also building ML-powered tooling that spots code similarities between new models and existing ones, with the eventual goal of auto-generating draft pull requests. Michael Han from Unsloth described the library as “the backbone of hundreds of thousands of projects,” and Wing Lian from Axolotl praised the standardized attention handling for simplifying post-training support.
Whether the community embraces the PyTorch-only future or fragments toward alternatives like JAX remains the open question. Hugging Face says it’s working with JAX ecosystem partners on compatibility, but the message is clear: in a field where reinvention is the price of relevance, v5 is a bet that doing less — and doing it cleaner — beats supporting everything.
💡 Key Takeaways
- Hugging Face is removing TensorFlow and Flax support entirely, betting its future on a PyTorch-only stack for Transformers.
- The library's model count has grown 10x since v4, forcing the team to build ML tools that auto-detect code similarities and automate contributions.
- Daily installs hit 3 million, but the team admits the maintenance burden of 400+ architectures required a modular rewrite just to stay viable.
- Unsloth and Axolotl maintainers publicly endorsed the changes, signaling that key downstream projects are aligned with the simplification push.
Keep reading: See related articles below for more coverage on this topic.
Get smarter about AI
The sharpest AI news, curated daily. Delivered free to your inbox.