Published: Apr 25, 2025

What Instella Means for the Future of Open AI Development

Open-source AI just got a serious upgrade. AMD has introduced Instella—a suite of 3B parameter language models trained entirely on AMD Instinct™ MI300X GPUs and released under a fully open framework. From training data to weights to the config files that shaped it, everything is out in the open.

At TensorWave, we’re not just excited—we’re aligned. Instella isn’t just a model drop. It’s a signal that open-source LLM development is moving beyond the walled gardens and into the hands of the builders. Here’s why that matters.

What Is Instella?

Instella is AMD’s new contribution to the open-source AI ecosystem:

  • 3B parameter models trained from scratch on MI300X
  • Released with model weights, training data sources, tokenizers, and PyTorch code
  • Available now via GitHub and Hugging Face

Performance-wise, Instella competes with some of the strongest open 3B-class models out there—Llama 3.2B, Qwen 2.5B, and Gemma 2B. And it was trained on the same infrastructure that’s powering the future of AI at scale: AMD Instinct MI300X.

Why This Drop Actually Matters

This isn’t just another open model—it’s a full-stack release:

  • Full transparency: From data mix to training scripts, devs get the full recipe—not just the result.
  • Inference- and fine-tune-ready: 3B parameters mean it’s deployable in production and adaptable with modest GPU clusters.
  • Built on ROCm + MI300X: It’s proof that AMD hardware isn’t just a CUDA alternative—it’s a competitive, open foundation for building.

Most so-called open models are trained on closed infra, then dumped with minimal context. Instella flips that narrative. It gives builders tools—not just weights.

What It Unlocks for Developers

For the open-source ML crowd, this is a goldmine:

  • Train-from-scratch experiments with full lineage
  • Remix datasets, re-tokenize, or re-architect
  • Fine-tune or deploy with zero vendor lock-in

Even more important: it gives developers a baseline model that was designed with openness and reproducibility from the start.

Run Instella Where It Was Born—On TensorWave

Instella was trained on MI300X. So naturally, it runs ridiculously well on MI300X—exactly what TensorWave delivers:

  • 🧠 256GB of HBM3 memory per GPU
  • ⚡ 5.3 TB/s memory bandwidth
  • 🛠️ ROCm-optimized, fine-tuned for inference and training alike

Want to spin it up? You’re minutes away:

  • Pre-built containers are live on TensorWave Cloud
  • Load from Hugging Face or run locally with our CLI tools
  • Persistent volume and caching options for lightning-fast iteration

Final Thoughts: AMD Is Building More Than Chips

With Instella, AMD isn’t just building silicon—they’re building momentum for open-source AI. They’re not just enabling training—they’re democratizing it.

At TensorWave, that’s our lane too. We’re here to make high-performance, open AI infrastructure accessible—whether you’re experimenting, scaling inference, or building something entirely new.

Instella is a step forward for the ecosystem. Let’s run it.

About TensorWave

TensorWave is the AI and HPC cloud purpose-built for performance. Powered exclusively by AMD Instinct™ Series GPUs, we deliver high-bandwidth, memory-optimized infrastructure that scales with your most demanding models—training or inference.

Ready to get started? Connect with a Sales Engineer.