Which AMD GPUs does Unsloth v0.1.50-beta support?

Unsloth v0.1.50-beta supports AMD Radeon RX 6000, 7000, 8000, and 9000 series consumer cards, the Instinct MI300X (192 GB) data center GPU, and Ryzen integrated graphics. It works on Windows, Windows Subsystem for Linux, and native Linux, and covers over 500 open models including Gemma 4, Qwen3.6, DeepSeek V4, Kimi K3, and GLM-5.2.

How does Unsloth on AMD compare to Unsloth on NVIDIA?

Unsloth reports the same headline gains on AMD as on NVIDIA: up to 2x faster training and roughly 70% less VRAM at full accuracy. The AMD path uses HIPified Triton kernels ported by the Unsloth team in collaboration with AMD's ROCm engineers, so training, reinforcement learning, chat, and deployment all work on AMD hardware.

Is Unsloth free and open source?

Yes. The core Unsloth package is Apache-2.0, and the optional Unsloth Studio UI is AGPL-3.0. Both are on GitHub at github.com/unslothai/unsloth, which has passed 68,000 stars, and there is no license fee for AMD or NVIDIA use.

What else changed in the v0.1.50-beta release?

Beyond AMD support, v0.1.50-beta adds a four-level tool-call permission selector (Ask, Approve for me, Off, Full access) for the agent runtime, an optional MCP endpoint that lets compatible AI clients inspect models and export GGUFs, and automatic HTTP retry for stalled Hugging Face downloads. Existing GGUF files are now reused to avoid redundant downloads.

Unsloth AI · 2026-07-20 · major

Unsloth v0.1.50-beta — AMD GPU support lands across Windows, WSL, and Linux

Unsloth v0.1.50-beta introduces AMD GPU support, enabling local fine-tuning and inference for 500+ models on Radeon, Instinct, Ryzen, and data center chips across Windows, WSL, and Linux — up to 2x faster with 70% less VRAM.

GitHub release page for Unsloth v0.1.50-beta introducing AMD support

Unsloth's LLM training and inference stack now runs natively on AMD GPUs across Windows, WSL, and Linux.

Key specs

GitHub stars	68.5K
Models supported	500+
Training speedup	up to 2x
Vram savings	~70%

Quick facts

Maker	Unsloth AI
Version	v0.1.50-beta (beta)
Released	July 20, 2026
License	Apache-2.0 (core); AGPL-3.0 (Studio UI)
AMD hardware	Radeon RX 6000–9000, Instinct MI300X, Ryzen, plus data center GPUs
Platforms	Windows, WSL, Linux
Speedup	Up to 2x faster training, ~70% less VRAM

What is it?

Unsloth v0.1.50-beta adds first-class AMD GPU support to the popular local training tool. Until now Unsloth's custom Triton kernels only ran on NVIDIA hardware; this release ports them to HIP/ROCm so Radeon, Instinct, and Ryzen GPUs can fine-tune and serve 500+ open models — Gemma 4, Qwen3.6, DeepSeek V4, Kimi K3, GLM-5.2, and more — on Windows, WSL, and Linux.

How does it work?

AMD support ships as HIPified Triton kernels tuned with the ROCm team for memory and profiling on AMD hardware. Unsloth v0.1.50-beta reports the same 2x training speed and 70% VRAM reduction it delivers on NVIDIA, and adds a four-level tool-call permission selector for its agent runtime, an optional MCP endpoint for external clients, and auto-retry for stalled Hugging Face downloads.

Why does it matter?

AMD hardware — especially the 192 GB Instinct MI300X and consumer Radeon cards — has been priced well below equivalent NVIDIA parts, but most fine-tuning tooling assumed CUDA. Unsloth landing AMD support in a release version, with Windows and WSL included, removes the last big blocker for tuning frontier open models on non-NVIDIA hardware and gives builders a real second GPU vendor for local work.

Who is it for?

developers fine-tuning open models on AMD hardware, self-hosters on tight VRAM budgets

Frequently asked questions

Which AMD GPUs does Unsloth v0.1.50-beta support?: Unsloth v0.1.50-beta supports AMD Radeon RX 6000, 7000, 8000, and 9000 series consumer cards, the Instinct MI300X (192 GB) data center GPU, and Ryzen integrated graphics. It works on Windows, Windows Subsystem for Linux, and native Linux, and covers over 500 open models including Gemma 4, Qwen3.6, DeepSeek V4, Kimi K3, and GLM-5.2.
How does Unsloth on AMD compare to Unsloth on NVIDIA?: Unsloth reports the same headline gains on AMD as on NVIDIA: up to 2x faster training and roughly 70% less VRAM at full accuracy. The AMD path uses HIPified Triton kernels ported by the Unsloth team in collaboration with AMD's ROCm engineers, so training, reinforcement learning, chat, and deployment all work on AMD hardware.
Is Unsloth free and open source?: Yes. The core Unsloth package is Apache-2.0, and the optional Unsloth Studio UI is AGPL-3.0. Both are on GitHub at github.com/unslothai/unsloth, which has passed 68,000 stars, and there is no license fee for AMD or NVIDIA use.
What else changed in the v0.1.50-beta release?: Beyond AMD support, v0.1.50-beta adds a four-level tool-call permission selector (Ask, Approve for me, Off, Full access) for the agent runtime, an optional MCP endpoint that lets compatible AI clients inspect models and export GGUFs, and automatic HTTP retry for stalled Hugging Face downloads. Existing GGUF files are now reused to avoid redundant downloads.

Try it

pip install unsloth==0.1.50b0