Graphcore and Hugging Face just expanded their IPU-optimized model library from a single BERT offering to ten ready-to-run transformers spanning NLP, speech, and computer vision. Each model ships with IPU configuration files and pre-trained or fine-tuned weights, which means developers skip the tedious porting work entirely.
The lineup is a who’s who of modern transformer architecture. On the NLP side: GPT-2 for text generation, RoBERTa and DeBERTa for masked language modeling and downstream fine-tuning, BART for summarization and translation, T5 with its text-to-text framework, and LXMERT for vision-language tasks. The speech stack includes Wav2Vec2 and HuBERT — the latter matching or beating wav2vec 2.0 on Librispeech and Libri-light benchmarks, even with as little as 10 minutes of fine-tuning data. Vision gets the ViT, which treats image patches like word tokens.
What’s notable here isn’t just the model count. Graphcore has been a Hugging Face Hardware Partner since 2021, and the first Optimum Graphcore model — BERT — only landed last year. Shipping nine more in that window suggests the IPU software stack has matured enough to handle diverse architectures without custom engineering for each one. That’s been the persistent knock against Graphcore: impressive silicon, but software that demanded too much expertise. This release is a direct counterargument.
The broader context: Nvidia’s CUDA moat is deep, but it’s not unbreachable. Every time a credible alternative makes popular models run with minimal friction, the switching cost drops. Ten models won’t dethrone the A100 overnight. But for teams already evaluating Graphcore’s IPUs on price-performance or memory bandwidth grounds, the Optimum library removes the biggest objection — that getting started takes weeks, not hours.
