December Release

[danielhanchen]

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

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

Unsloth Zoo Changes

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

New Contributors

• @dmsuehir made their first contribution in Fix llama tokenizer padding_side when using model.generate in inference mode #3644
• @gitpullpull made their first contribution in Fix broken link for advanced pip installation in README #3652
• @bhuvanprakash made their first contribution in Fix: prevent load_in_fp8 kwarg from reaching Qwen3MoeForCausalLM constructor (Fix #3649) #3654
• @ved1beta made their first contribution in FIX set defualt [128, 128] insted of none #3658
• @sbhavani made their first contribution in Fix: Pass gradient_checkpointing parameter to model.for_training() #3659
• @noah1510 made their first contribution in Update transformers version constraint in pyproject.toml #3689
• @ChetanKrishna07 made their first contribution in Update e to error unsloth-zoo#374
• @electroglyph made their first contribution in update compiler for XLMRobertaModel unsloth-zoo#383

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

---

This discussion was created from the release December Release.