v0.4.0

What's Changed

• Improve tic-tac-toe self-play recipe by @YujiaBao in #634
• Extract tic-tac-toe opponents from env.py into tictactoe.py (#634 follow-up) by @YujiaBao in #647
• Fix broken quickstart link in README by @YujiaBao in #649
• Fix 8 broken tinker-docs links by @YujiaBao in #650
• Consolidate 7 skills into 2 (research + debug) and fix plugin manifest by @YujiaBao in #651
• Fix broken RL Environments link in tutorial 301 by @YujiaBao in #657
• Bump transformers upper bound to 5.5.3, drop CI version matrix by @YujiaBao in #654
• tutorials: add an optional text input for the Tinker API key by @akshayka in #656
• fix(docs): update TokensWithLogprobs usage in tutorial to use maybe_logprobs by @tyfeng1997 in #658
• fix: normalize SFT weights per-example for consistent gradient magnitudes by @bledden in #334
• Add more runahead to supervised training loop by @sshleifer in #648
• feat: add Nemotron-3 low thinking renderer for Super model by @YujiaBao in #667
• fix(docs): correct TinkerAPI to Tinker API in recipes README by @Jah-yee in #669
• feat: add support for Qwen3.6-35B-A3B by @YujiaBao in #674
• feat: add Kimi-K2.6 model support by @YujiaBao in #675
• sdft: reverse KL fix by @likenneth in #677
• feat: add support for Qwen3.6-27B by @YujiaBao in #681
• Unify checkpoint machinery by @danobi in #676
• fix(pyright): disable reportPrivateImportUsage for torch by @YujiaBao in #686
• [ckpt] Support fire-and-forget periodic checkpoints by @danobi in #687
• Fix failing CI tests by @nealwu in #691
• Updates for 101 Hello Tinker by @nealwu in #690
• Replace parse_success bool with ParseTermination enum (fixes base-model eval regression) by @dphuang2 in #688
• HarborReward._upload_tests: recurse into nested fixture dirs by @danqi in #693
• evaluate_task: catch sandbox creation errors by @danqi in #694
• Add cross-account checkpoint copy recipe by @derek-tml in #692
• Empty commit to re-run CI by @nealwu in #696
• Align Kimi K2.6 revision pin by @derek-tml in #697

New Contributors

• @akshayka made their first contribution in #656
• @tyfeng1997 made their first contribution in #658
• @sshleifer made their first contribution in #648
• @Jah-yee made their first contribution in #669
• @likenneth made their first contribution in #677
• @danqi made their first contribution in #693
• @derek-tml made their first contribution in #692

Full Changelog: v0.3.0...v0.4.0