AMD Just Proved You Don't Need Nvidia to Train a Competitive AI Model

AMD has just delivered one of the clearest challenges yet to the idea that serious AI training only happens on Nvidia hardware. Its fully open Instella-3B model family was trained from scratch on AMD's own Instinct GPUs — and on November 8, 2026, it showed up on Huawei's Ascend ecosystem hub, Modelers.cn, putting a mainstream Western open model directly in front of Ascend NPU users.

What Instella-3B Actually Is

Instella-3B is a family of 3-billion-parameter language models that AMD released as fully open: the weights, the training code, the data recipe and even intermediate checkpoints are all public under the Apache 2.0 licence. The family covers several flavours — the raw base model for text completion, the instruction-tuned Instella-3B-SFT and DPO-tuned Instella-3B-Instruct for chat, Instella-Long with a 128K context window, and Instella-Math for reasoning. It joins a growing wave of open-weight models that anyone can download, inspect and run.

The training story is the point. AMD built Instella-3B from scratch on 128 of its own MI300X Instinct GPUs across 16 nodes, as described in its paper (arXiv 2511.10628, November 2025). Stage 1 covered roughly 4.07 trillion tokens, followed by a second stage of 57 billion tokens focused on reasoning-heavy domains. AMD's own evaluations put the result in the same league as other small open models such as Qwen2.5-3B, Llama 3.2 3B, Gemma 2 2B and OLMo 2.

Why the Huawei Listing Matters

On November 8, 2026, Instella-3B-SFT and the Stage 1 checkpoint appeared on Modelers.cn, Huawei's hub for the Ascend ecosystem. That gives Ascend NPU users direct access to a mainstream Western open model, at around 6-7GB in bf16 — light enough to run locally. Two honest caveats: the listing itself is sparse on documentation, and its licence field reads 'other', even though the upstream AMD release is Apache 2.0. Users should check the terms before building on it.

The Bigger Picture

This matters for three reasons. First, hardware diversity: Instella-3B is proof that competitive models can be trained end to end without Nvidia silicon. Second, transparency: because the code, data recipe and checkpoints are public, researchers can reproduce and audit the work instead of taking it on trust — a win for open-source AI. Third, timing: capable new models are arriving fast from every direction, and Tencent Hy4 is another recent example. The assumption that one vendor owns AI training is looking shakier by the month.

Our Take

Instella-3B is not the biggest model of the year, and at 3 billion parameters it is not trying to be. Its significance is structural: an open model, trained on non-Nvidia hardware, now running on Huawei's Ascend platform. The hardware monopoly is cracking — and open models are doing the cracking.