AMD released its own open model, Instella. The detail that makes it news isn't the performance — it's where it was trained.
Nvidia dominates the sector because nearly every AI tool is built on the CUDA platform, and training outside of it is notoriously difficult. AMD trained Instella from scratch on AMD Instinct chips, using the ROCm software stack.
Why this matters
- It's practical proof that you can train a competitive model outside the CUDA ecosystem.
- AMD published the checkpoints for each stage and the training recipes, not just the final model.
- According to the source, it outperforms models of similar size, like Gemma 4 E4B and a smaller version of Qwen 3.5.
1.Why 'without CUDA' is the headline
2 minNvidia basically took over the AI space because most tools are built on the CUDA platform. It's very hard to train and run models on chips that aren't CUDA.
A single supplier at the most expensive point in the chain is a structural problem: it sets the price, delivery time, and who gets to train. A competitive model trained entirely on AMD hardware is proof that a second path exists — and that matters more than where Instella lands on any performance chart.
Spec sheet
- License
- Open source (code, weights, and recipes)
- Parameters
- 16 B totais / 2,8 B ativos
- Trained on
- AMD Instinct + pilha ROCm
Continue in the full microcourse
You've read the opening of 3 classes
The microcourse covers the complete step-by-step, the selection criteria, where the tool fails, who it's really for — and, in the Expert version, the official address to start today.
- The spec sheet and the two techniques2 min
- What AMD published alongside it3 min
How we verified
We track releases straight from primary sources, transcribe what's demonstrated, check every name and number against the manufacturer's official documentation, and rewrite it in Portuguese — with what the tool no do it together, which is the part the ad leaves out.
