unslothai/unsloth December-2025

December Release + 3x Faster Training

on GitHub

Thanks for all the love and support this year! We're wishing you all a lovely Christmas. Please update Unsloth & our Docker to use the latest updates! 🦥

• Introducing 3x faster training & 30% less VRAM. New Triton kernels, padding-free & packing. Blog
• 500K Context training and reinforcement learning is now possible on a single 80GB GPU. Blog • Notebook
• Fine-tune then Deploy LLMs on your Phone with PyTorch and Unsloth. Tweet • Read Guide
• 🤗 Transformers v5 is now supported! It's not enabled by default due to possible instability issues.
• Preliminary multi-GPU support: DDP Guide (not representative of the official release early next year)
• More: Sudoku RL nb • Paddle-OCR nb • New NVIDIA blog
• Lots of bug fixes! See further below.

🔮 New Models + Guides

• ✨FunctionGemma: Google new 270M tool-calling LLM. Guide • Notebook
• Nemotron 3: NVIDIA new 30B reasoning model. Guide • GGUF
• Mistral: new coding & instruct VLMs. Ministral 3 • Devstral 2
• GLM-4.6V: new vision models. Guide • 4.6V • 4.6V-Flash
• More: Qwen3-Next • Mistral Large 3 • FLUX.2-dev

Bug Fixes and Enhancements

1. Supports rollout_func allowing multi turn RL to work
2. Supports vllm>=0.12.0 and efficient GRPO for it
3. Supports transformers>=5.0.0, first shown via our Ministral notebooks
4. Fix HuggingFace token logins not working for private repos
5. Fixes TorchAO and QAT not working during saving
6. Fixed DeepSeek OCR finetuning not loading finetuned models
7. Improved vision utilities for vision VLM finetuning

What's Changed

• Fix llama tokenizer padding_side when using model.generate in inference mode by @dmsuehir in #3644
• Fix indefinite article usage in comments and docstrings by @mk0walsk in #3648
• fix rope_theta -> rope_parameters['rope_theta'] by @mmathew23 in #3651
• Fix broken link for advanced pip installation in README by @gitpullpull in #3652
• Fix: prevent load_in_fp8 kwarg from reaching Qwen3MoeForCausalLM constructor (Fix #3649) by @bhuvanprakash in #3654
• make unsloth_tiled_mlp a from_pretrained arg by @mmathew23 in #3655
• FIX set defualt [128, 128] insted of none by @ved1beta in #3658
• Fix: Pass gradient_checkpointing parameter to model.for_training() by @sbhavani in #3659
• [FIX] Vllm guided decoding params by @Datta0 in #3662
• Vllm guided decoding by @Datta0 in #3663
• Nightly by @danielhanchen in #3664
• [pre-commit.ci] pre-commit autoupdate by @pre-commit-ci[bot] in #3666
• Update transformers version constraint in pyproject.toml by @noah1510 in #3689
• [pre-commit.ci] pre-commit autoupdate by @pre-commit-ci[bot] in #3694
• Remove reload_weights rpc call from grpo trainer by @Datta0 in #3673
• [Fix] [TRL] load_lora for multi line llm.chat/generate by @Datta0 in #3696
• Nightly by @danielhanchen in #3698
• SFT sample packing by @djsaunde in #3566
• Auto-enable padding-free SFT by @djsaunde in #3672
• [FIX] fbgemm version check by @Datta0 in #3704
• Nightly by @danielhanchen in #3706
• update TRL filter by @djsaunde in #3707
• [intel] skip xpu fbgemm fp8 by @leizhenyuan in #3625
• Mistral packing, train on completions only, simplifications by @djsaunde in #3709
• Update torchao save by @metascroy in #3679
• Nightly by @danielhanchen in #3720
• [pre-commit.ci] pre-commit autoupdate by @pre-commit-ci[bot] in #3731
• Bug fixes by @danielhanchen in #3734
• Update FUNDING.yml by @danielhanchen in #3736
• Nightly by @danielhanchen in #3737
• Fix Deepseek OCR Lora Model Load by @mmathew23 in #3738

Unsloth Zoo Changes

• updates for vLLM compativility with lora by @danielhanchen in unslothai/unsloth-zoo#359
• Nightly by @danielhanchen in unslothai/unsloth-zoo#355
• Add logging to tiled mlp and fix target chunk size calculation by @mmathew23 in unslothai/unsloth-zoo#361
• Remove include_buffers from init_empty_weights by @pluesclues in unslothai/unsloth-zoo#363
• packed seq lengths token count correction by @djsaunde in unslothai/unsloth-zoo#348
• Configure ce target gb by @mmathew23 in unslothai/unsloth-zoo#365
• [FIX] vLLM LoRA extra vocab by @Datta0 in unslothai/unsloth-zoo#367
• Nightly by @danielhanchen in unslothai/unsloth-zoo#368
• [FIX] vLLM local lora tensor loading by @Datta0 in unslothai/unsloth-zoo#370
• vllm lora_dir rename and make embedding padding optional by @danielhanchen in unslothai/unsloth-zoo#373
• Bug fixes by @danielhanchen in unslothai/unsloth-zoo#375
• Update e to error by @ChetanKrishna07 in unslothai/unsloth-zoo#374
• Vision utils decode image improvement by @mmathew23 in unslothai/unsloth-zoo#372
• [FIX] [DDP] Fix compile for distributed training by @Datta0 in unslothai/unsloth-zoo#379
• Nightly by @danielhanchen in unslothai/unsloth-zoo#382
• update compiler for XLMRobertaModel by @electroglyph in unslothai/unsloth-zoo#383
• Fix Deepseek OCR Lora Model Load by @mmathew23 in unslothai/unsloth-zoo#386
• fix for non-generation models in transformers 5 by @electroglyph in unslothai/unsloth-zoo#388

New Contributors

• @dmsuehir made their first contribution in #3644
• @gitpullpull made their first contribution in #3652
• @bhuvanprakash made their first contribution in #3654
• @ved1beta made their first contribution in #3658
• @sbhavani made their first contribution in #3659
• @noah1510 made their first contribution in #3689
• @ChetanKrishna07 made their first contribution in unslothai/unsloth-zoo#374
• @electroglyph made their first contribution in unslothai/unsloth-zoo#383

Full Changelog: November-2025...December-2025