December-2025

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! 🦥<br>
<img width="400" height="1020" alt="Unsloth December Release" src="https://github.com/user-attachments/assets/bc204ecf-6b28-4ab5-aa7f-c98de2eb7e0a" />
* Introducing **3x faster training** & 30% less VRAM. New Triton kernels, padding-free & packing. [Blog](https://docs.unsloth.ai/new/3x-faster-training-packing)
* **500K Context** training and reinforcement learning is now possible on a single 80GB GPU. [Blog](https://docs.unsloth.ai/new/500k-context-length-fine-tuning) • [Notebook](https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/gpt_oss_(20B)_500K_Context_Fine_tuning.ipynb)
* Fine-tune then Deploy LLMs on your **Phone** with PyTorch and Unsloth. [Tweet](https://x.com/UnslothAI/status/2001305185206091917) • [Read Guide](https://docs.unsloth.ai/new/deploy-llms-phone)
* 🤗 Transformers v5 is now supported! It's not enabled by default due to possible instability issues.
* Preliminary **multi-GPU support**: [DDP Guide](https://docs.unsloth.ai/basics/multi-gpu-training-with-unsloth/ddp) (not representative of the official release early next year)
* More: [Sudoku RL nb](https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/Ministral_3_(3B)_Reinforcement_Learning_Sudoku_Game.ipynb) • [Paddle-OCR nb](https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/Paddle_OCR_(1B)_Vision.ipynb) • [New NVIDIA blog](https://blogs.nvidia.com/blog/rtx-ai-garage-fine-tuning-unsloth-dgx-spark/)
* Lots of bug fixes! See further below.

### :crystal_ball: __New Models + Guides__
* **:sparkles:FunctionGemma**: Google new 270M tool-calling LLM. [Guide](https://docs.unsloth.ai/models/functiongemma) • [Notebook](https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/FunctionGemma_(270M).ipynb)
* **Nemotron 3**: NVIDIA new 30B reasoning model. [Guide](https://docs.unsloth.ai/models/nemotron-3) • [GGUF](https://huggingface.co/unsloth/Nemotron-3-Nano-30B-A3B-GGUF)
* **Mistral**: new coding & instruct VLMs. [Ministral 3](https://docs.unsloth.ai/models/ministral-3) • [Devstral 2](https://docs.unsloth.ai/models/devstral-2)
* **GLM-4.6V**: new vision models. [Guide](https://docs.unsloth.ai/models/glm-4.6-how-to-run-locally) • [4.6V](https://huggingface.co/unsloth/GLM-4.6V-GGUF) • [4.6V-Flash](https://huggingface.co/unsloth/GLM-4.6V-Flash-GGUF)
* More: [Qwen3-Next](https://docs.unsloth.ai/models/qwen3-next) • [Mistral Large 3](https://huggingface.co/unsloth/Mistral-Large-3-675B-Instruct-2512-GGUF) • [FLUX.2-dev](https://huggingface.co/unsloth/FLUX.2-dev-GGUF) 

> [!TIP]
> Update Unsloth via `pip install --upgrade --force-reinstall --no-cache-dir --no-deps unsloth unsloth_zoo`
> If you want PyTorch 2.9: `pip install --upgrade unsloth unsloth_zoo`

## 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
4. Supports `transformers>=5.0.0`, first shown via our Ministral notebooks
5. Fix HuggingFace token logins not working for private repos
6. Fixes TorchAO and QAT not working during saving
7. Fixed DeepSeek OCR finetuning not loading finetuned models
8. 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 https://github.com/unslothai/unsloth/pull/3644
* Fix indefinite article usage in comments and docstrings by @mk0walsk in https://github.com/unslothai/unsloth/pull/3648
* fix rope_theta -> rope_parameters['rope_theta'] by @mmathew23 in https://github.com/unslothai/unsloth/pull/3651
* Fix broken link for advanced pip installation in README by @gitpullpull in https://github.com/unslothai/unsloth/pull/3652
* Fix: prevent load_in_fp8 kwarg from reaching Qwen3MoeForCausalLM constructor (Fix #3649) by @bhuvanprakash in https://github.com/unslothai/unsloth/pull/3654
* make unsloth_tiled_mlp a from_pretrained arg by @mmathew23 in https://github.com/unslothai/unsloth/pull/3655
* FIX set defualt [128, 128]  insted of none by @ved1beta in https://github.com/unslothai/unsloth/pull/3658
* Fix: Pass gradient_checkpointing parameter to model.for_training() by @sbhavani in https://github.com/unslothai/unsloth/pull/3659
* [FIX] Vllm guided decoding params by @Datta0 in https://github.com/unslothai/unsloth/pull/3662
* Vllm guided decoding by @Datta0 in https://github.com/unslothai/unsloth/pull/3663
* Nightly by @danielhanchen in https://github.com/unslothai/unsloth/pull/3664
* [pre-commit.ci] pre-commit autoupdate by @pre-commit-ci[bot] in https://github.com/unslothai/unsloth/pull/3666
* Update transformers version constraint in pyproject.toml by @noah1510 in https://github.com/unslothai/unsloth/pull/3689
* [pre-commit.ci] pre-commit autoupdate by @pre-commit-ci[bot] in https://github.com/unslothai/unsloth/pull/3694
* Remove reload_weights rpc call from grpo trainer by @Datta0 in https://github.com/unslothai/unsloth/pull/3673
* [Fix] [TRL] load_lora for multi line llm.chat/generate by @Datta0 in https://github.com/unslothai/unsloth/pull/3696
* Nightly by @danielhanchen in https://github.com/unslothai/unsloth/pull/3698
* SFT sample packing by @djsaunde in https://github.com/unslothai/unsloth/pull/3566
* Auto-enable padding-free SFT by @djsaunde in https://github.com/unslothai/unsloth/pull/3672
* [FIX] fbgemm version check by @Datta0 in https://github.com/unslothai/unsloth/pull/3704
* Nightly by @danielhanchen in https://github.com/unslothai/unsloth/pull/3706
* update TRL filter by @djsaunde in https://github.com/unslothai/unsloth/pull/3707
* [intel] skip xpu fbgemm fp8 by @leizhenyuan in https://github.com/unslothai/unsloth/pull/3625
* Mistral packing, train on completions only, simplifications by @djsaunde in https://github.com/unslothai/unsloth/pull/3709
* Update torchao save by @metascroy in https://github.com/unslothai/unsloth/pull/3679
* Nightly by @danielhanchen in https://github.com/unslothai/unsloth/pull/3720
* [pre-commit.ci] pre-commit autoupdate by @pre-commit-ci[bot] in https://github.com/unslothai/unsloth/pull/3731
* Bug fixes by @danielhanchen in https://github.com/unslothai/unsloth/pull/3734
* Update FUNDING.yml by @danielhanchen in https://github.com/unslothai/unsloth/pull/3736
* Nightly by @danielhanchen in https://github.com/unslothai/unsloth/pull/3737
* Fix Deepseek OCR Lora Model Load by @mmathew23 in https://github.com/unslothai/unsloth/pull/3738

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

## New Contributors
* @dmsuehir made their first contribution in https://github.com/unslothai/unsloth/pull/3644
* @gitpullpull made their first contribution in https://github.com/unslothai/unsloth/pull/3652
* @bhuvanprakash made their first contribution in https://github.com/unslothai/unsloth/pull/3654
* @ved1beta made their first contribution in https://github.com/unslothai/unsloth/pull/3658
* @sbhavani made their first contribution in https://github.com/unslothai/unsloth/pull/3659
* @noah1510 made their first contribution in https://github.com/unslothai/unsloth/pull/3689
* @ChetanKrishna07 made their first contribution in https://github.com/unslothai/unsloth-zoo/pull/374
* @electroglyph made their first contribution in https://github.com/unslothai/unsloth-zoo/pull/383

**Full Changelog**: https://github.com/unslothai/unsloth/compare/November-2025...December-2025