Burn v0.22.0-pre.4 and CubeCL v0.11.0-pre.4: paired Tracel pre-releases

On 2026-09-22, Tracel tagged paired pre-releases: Burn v0.22.0-pre.4 and CubeCL v0.11.0-pre.4. Both GitHub releases are marked Pre-release—this is not stable Burn 0.22 or CubeCL 0.11.

This post is based on the release notes and the project READMEs. Burn and CubeCL are dual licensed MIT / Apache-2.0, as stated on their READMEs. It gives no speedup or GFLOPS figures of its own.

What they are

Burn is Tracel’s Rust tensor library and deep-learning framework. CubeCL is the multi-platform compute language / JIT / runtime behind Burn’s accelerated backends, and is usable standalone. CubeCL’s README still calls the project alpha, with a public API that can break between minor versions.

Burn highlights (from the pre.4 body)

Changelog items from the Burn tag include:

• LAMB optimizer
• einsum with runtime and macro APIs
• Breaking extracts of tensor linalg and signal into burn-linalg / burn-signal extension crates
• Merge of AutodiffModule into Module; explicit autodiff conversions replacing no_grad; separate gradient control from module freezing
• Asymmetric padding in conv1d / conv2d
• Deprecation of the burn-tch (LibTorch) backend on the 0.22 train (no removal date stated here)
• Docs / migration guides for Burn 0.22 (#5762, #5768)
• Store/interop and dataset work (PyTorch reader extract, atomic safetensors writes, SqliteDataset on Turso)

Several entries carry !: breaking markers—treat migration cost as real and follow Tracel’s 0.22 migration docs rather than inventing steps.

CubeCL highlights (from the pre.4 body)

The CubeCL tag (published ~5 minutes before Burn) centers on backends and tooling:

• AMDGPU backend (gated off macOS / default features in follow-ons)
• CUDA via LLVM backend and related LLVM / PTX / NVPTX fixes
• Device timing / profiling for HIP and CUDA
• IR and memory work (MemorySSA, uniformity analysis, physical card reporting, complex support, dp4a)
• CPU vectorized math approximations and layout / runtime refactors

Who should care

Rust ML teams already on the Burn / CubeCL pre-release train who need the latest optimizer, tensor-API, and backend work should read both tags. Everyone else should wait for stable 0.22 / 0.11 unless they are deliberately tracking pre-releases.