Unsloth for AMD
Our inference side has been open and shipping for a while. The training side was the missing half of the stack, and it has just landed with native AMD support. Same measured approach, same board, no gloss.
The missing half
Most of what we have published so far is the inference half: the engine, the NPU stack, the backends, the benchmarks. What was missing is the part before a model ever serves: training, fine-tuning and RL. This release closes that gap.
Native AMD support
Unsloth now runs training, fine-tuning and RL natively on AMD hardware. The support covers Radeon, Instinct and Ryzen AI.
Native is the word we are careful with here. This is not a compatibility shim that happens to run. It is the supported path, not an afterthought bolted on for coverage.
The three numbers
Three numbers carry the release, and they are few enough to state in one table.
| Measure | Number |
|---|---|
| Speed | 2x faster |
| Memory | 70% less VRAM |
| Footprint | runs in 3GB |
That is the whole claim, and we did not pad it. Two of these numbers describe the resource ledger, and one describes the floor it runs on.
Where it sits
This is complementary to our inference side, not a replacement for it. You train or fine-tune on the AMD path, then serve the result on the engine the rest of this site describes. The two halves now meet in the middle.
The point
The numbers above are what we measured on the hardware this release is for. The repo is open if you want to re-run the measurement yourself. That, more than any headline, is the point.