Download Latest Version tinygrad 0.14.0 source code.zip (12.7 MB)
Email in envelope

Get an email when there's a new version of tinygrad

Home / v0.14.0
Name Modified Size InfoDownloads / Week
Parent folder
README.md 2026-08-24 1.1 kB
0
tinygrad 0.14.0 source code.tar.gz 2026-08-24 11.9 MB
0
tinygrad 0.14.0 source code.zip 2026-08-24 12.7 MB
0
Totals: 3 Items   24.6 MB 0

chestnut release! get your chestnut here: https://comma.ai/shop/chestnut Over 1000 commits since 0.13.0. At 26040 lines.

Release Highlights

  • tinygrad.llm improvements
  • RDNA3 kernels for Qwen 3.8 27B [#17512]
  • Faster tokenizer [#17077]
  • Tool calling support [#17061]
  • USB interface has faster copy-in [#17628] [#17663]
  • Tensor.training and Tensor.train() removed in favour of TRAINING ContextVar [#16703] [#16705] [#16774]
  • Python scalars are now weak dtypes (dtypes.weakint/dtypes.weakfloat) [PRs]
  • DEFAULT_FLOAT and DEFAULT_INT are ContextVars [#17265]
  • Kernels are now compiled in parallel by default [#17629]
  • llama3.1 8B training improvements [PRs]
  • GPT-OSS 20B added as a mlperf training benchmark [#16837] [#16884]

See the full changelog: https://github.com/tinygrad/tinygrad/compare/v0.13.0...v0.14.0

Join the Discord!

Source: README.md, updated 2026-08-24