Burn 0.22 and CubeCL 0.11 land as the stable tags

On 2026-10-06 Tracel tagged the paired stables Burn v0.22.0 (19:25:58Z) and CubeCL v0.11.0 (18:45:19Z). crates.io lists burn 0.22.0 (18:46:00Z) and cubecl 0.11.0 (15:59:12Z). Neither crate version is yanked. In this brief, stable means that pair: a GitHub tag with no -pre suffix, plus the crates.io release. It does not mean CubeCL has left alpha.

Follow-up to Burn v0.22.0-pre.4 and CubeCL v0.11.0-pre.4, which told everyone outside the pre-release train to wait for these tags.

The CubeCL README at v0.11.0 still says the project is in alpha, that the public API can break between minor versions, and that direct dependents should pin a version. Burn and CubeCL remain dual-licensed MIT or Apache-2.0, as that README states for CubeCL.

What the Burn release body says

From the v0.22.0 summary and migration pointer:

• Models and tensors no longer carry a backend generic. User code picks a Device (Device::cuda(0), Device::wgpu(..), Device::flex()). The Backend trait stays for implementations. Training turns autodiff on with device.autodiff() before the model and its inputs are initialized.
• LoRA and QLoRA are in the release (Feat/lora qlora, #5139).
• Graph capture is in the release (Feat/graph capture, #5146; a capture backend that produces GraphIr, #5377). The Tracel blog describes replay on CUDA and HIP, and a software graph for WGPU replay. That is Tracel’s account of the mechanism.
• Burn says Pliron replaces the MLIR-based intermediate representation, and Turso replaces the bundled SQLite dependency. The changelog entries are the Pliron migration (#5324) and SqliteDataset on Turso instead of rusqlite (#5546).
• Burn says CubeCL adds LLVM-based CUDA and AMD GPU compilation paths. LLVM stays a native dependency for the backends that use it. The CubeCL tag includes an AMDGPU backend (#1571) and a CUDA LLVM backend (#1609).

The migration guide puts the minimum supported Rust version at 1.95 and says burn has no default execution backend. Enable a feature for each device constructor you call.

Candle is removed. NdArray and LibTorch are not

The release body says Candle has been removed in this release, and that NdArray and LibTorch are being deprecated. The changelog splits them the same way:

• Removed: chore!: remove deprecated burn-candle backend (#5343). The migration guide maps the candle, candle-cuda, and candle-metal features to Removed.
• Deprecated: chore: deprecate burn-ndarray backend (#5344) and chore: deprecate burn-tch backend (#5525).

The burn crate text on crates.io says the LibTorch backend is deprecated as of 0.22.0 and will be removed in a future release. It names no date. NdArray’s deprecation note in the release does not give a removal date either.

Benches

The Tracel blog reports faster model-edit rebuilds and training-step changes on the authors’ machines. Those figures are self-reported. This brief does not repeat them.

Rust ML teams that held after the pre.4 post should read Migrating to Burn 0.22 before bumping. Anyone depending on CubeCL directly should still pin the crate: the v0.11.0 README still expects breaking changes between minors.