✨ Gemma 3n now available

• Google's new Gemma 3n multimodal models that support text, image, video & audio. Guide
• Gemma 3n finetuning notebook + audio, vision, text inference Colab notebook
• Gemma 3n collection in dynamic GGUF, safetensor 4-bit etc formats: Gemma-3n

🎵 Text-to-Speech (TTS) Fine-tuning

• Train TTS/STT models like Sesame-CSM, Orpheus-TTS and OpenAI's Whisper locally! Guide
• Clone voices, learn new emotions, tones & styles with 1.5x faster training and -50% VRAM. Notebooks

[!TIP] Update Unsloth via pip install --upgrade --force-reinstall unsloth unsloth_zoo

🧠 DeepSeek-R1-0528 Support with Dynamic 1-bit GGUFs

• Fine-tune DeepSeek-R1-0528-Qwen3 with GRPO! Our new reward function increases multilingual response rates by 40%+ Notebook
• Dynamic 1-bit GGUFs shrink the full 715GB model to just 175GB (-80% size)

📈 Dynamic 2.0 GGUFs

• New quantization method that achieves SOTA performance. More info
• Sets new benchmarks for 5-shot MMLU and KL Divergence and selectively quantizes layers for optimal accuracy

⚡ Advanced Qwen3 GRPO notebook

• Proximity scoring for more better reward functions. Advanced GRPO notebook
• New Prefinetuning/priming to skip GRPO format learning

🎯 Magistral Conversational Reasoning

• Fine-tune Magistral-24B for advanced conversational reasoning. Notebook

👁️ Gemma3 Vision Support

• Fine-tune Gemma3 vision models for multimodal tasks Notebook

Documentation & Guides

• Reinforcement Learning Guide: Complete guide on RL for LLMs covering GRPO, RLHF, DPO. Guide
• LoRA Hyperparameters Guide: Master optimal learning rates, epochs, LoRA rank & alpha settings. Guide

What's Changed

• Nightly by @danielhanchen in https://github.com/unslothai/unsloth/pull/2448
• Added k_norm & q_norm to merged Qwen3 layers by @cblomert in https://github.com/unslothai/unsloth/pull/2452
• MoE Kernel by @jeromeku in https://github.com/unslothai/unsloth/pull/2465
• Blackwell Support by @johnnynunez in https://github.com/unslothai/unsloth/pull/2458
• Added missing code of conduct by @rolandtannous in https://github.com/unslothai/unsloth/pull/2416
• Fix readme example by @yuanzhedong in https://github.com/unslothai/unsloth/pull/2492
• the pixtral vision notebook fails during inference by @mmathew23 in https://github.com/unslothai/unsloth/pull/2466
• [1/N] Enable intel GPU for unsloth by @leizhenyuan in https://github.com/unslothai/unsloth/pull/2350
• [2/N] Enable intel GPU for unsloth by @leizhenyuan in https://github.com/unslothai/unsloth/pull/2388
• vLLM Windows CUDA support [tested] by @fenglui in https://github.com/unslothai/unsloth/pull/2158
• Add Sesame CSM by @mmathew23 in https://github.com/unslothai/unsloth/pull/2527
• Add Qwen-3 chat template and Ollama template support by @kiankyars in https://github.com/unslothai/unsloth/pull/2537
• Fix typos by @omahs in https://github.com/unslothai/unsloth/pull/2540
• Add use_rslora reference to LoraConfig inititalisation by @jkumz in https://github.com/unslothai/unsloth/pull/2539
• TTS by @danielhanchen in https://github.com/unslothai/unsloth/pull/2545
• Quick fix on the CompileConfig error by @Erland366 in https://github.com/unslothai/unsloth/pull/2554
• Fix trust remote code by @Etherll in https://github.com/unslothai/unsloth/pull/2357
• fix issue with qwen3 template double quote escapes by @davedgd in https://github.com/unslothai/unsloth/pull/2563
• Display the model name in RoPE scaling unsupported error by @emmanuel-ferdman in https://github.com/unslothai/unsloth/pull/2564
• Fix Whisper, ModernBERT by @danielhanchen in https://github.com/unslothai/unsloth/pull/2565
• fix: improved error handling when llama.cpp build fails #2358 by @Hansehart in https://github.com/unslothai/unsloth/pull/2603
• Remove dataset_text_field from SFTConfig by @qgallouedec in https://github.com/unslothai/unsloth/pull/2609
• Upgrade trl fix by @Datta0 in https://github.com/unslothai/unsloth/pull/2544
• Check the skip_prepare_dataset before accessing dataset fields. #2496 by @Premik in https://github.com/unslothai/unsloth/pull/2633
• Llama4 MoE Grouped GEMM by @jeromeku in https://github.com/unslothai/unsloth/pull/2639
• Latest TRL, GRPO + Bug fixes by @danielhanchen in https://github.com/unslothai/unsloth/pull/2645
• Fix SFTtraining for new trl by @mmathew23 in https://github.com/unslothai/unsloth/pull/2647
• Bug fixes by @danielhanchen in https://github.com/unslothai/unsloth/pull/2651
• Fix quant model param fetch regex by @Datta0 in https://github.com/unslothai/unsloth/pull/2662
• Fix batched generation for prompts of different lengths by @RunFMe in https://github.com/unslothai/unsloth/pull/2216
• reroute merge logic language models + comprehensive tests + eval kits by @rolandtannous in https://github.com/unslothai/unsloth/pull/2673
• unsloth checkpointing fix for latest transformers==4.52.x by @mmathew23 in https://github.com/unslothai/unsloth/pull/2674
• patch sft_trainer to favor max_seq_length over max_length in config by @mmathew23 in https://github.com/unslothai/unsloth/pull/2669
• Update prepare 4d causal attention call by @mmathew23 in https://github.com/unslothai/unsloth/pull/2678
• Ignore None Values when building vllm subprocess_command by @Salpingopharyngeus in https://github.com/unslothai/unsloth/pull/2680
• add support for torch270 with Intel GPU by @leizhenyuan in https://github.com/unslothai/unsloth/pull/2709
• Making protobuf version more flexible by @user799595 in https://github.com/unslothai/unsloth/pull/2637
• tests for additional merge fix unsloth zoo pr 163 by @rolandtannous in https://github.com/unslothai/unsloth/pull/2719
• Reward modeling update (There seems to be another patch) by @pluesclues in https://github.com/unslothai/unsloth/pull/2710
• Fix Typos in Documentation and Comments by @leopardracer in https://github.com/unslothai/unsloth/pull/2721
• Fix renaming on other model than Llama by @Erland366 in https://github.com/unslothai/unsloth/pull/2762
• Enable vLLM to share memory space by @Datta0 in https://github.com/unslothai/unsloth/pull/2712
• Fix TRL 1.8.2 by @marcandrelarochelle in https://github.com/unslothai/unsloth/pull/2774
• Fix AttributeError in GRPO trainer for models without llm attribute by @rolandtannous in https://github.com/unslothai/unsloth/pull/2780
• Additional tests for unsloth-zoo PR#174 by @rolandtannous in https://github.com/unslothai/unsloth/pull/2779
• Update pyproject.toml by @amrothemich in https://github.com/unslothai/unsloth/pull/2778
• Fix for grpo_compute_loss_slow by @simpissa in https://github.com/unslothai/unsloth/pull/2702
• Fix GRPO by @danielhanchen in https://github.com/unslothai/unsloth/pull/2787
• Docs: Fix typo and improve MoE docstrings by @kilavvy in https://github.com/unslothai/unsloth/pull/2784
• [5/N] Enable intel GPU for unsloth by @leizhenyuan in https://github.com/unslothai/unsloth/pull/2768
• Sequence Classification Bug Fixes by @pluesclues in https://github.com/unslothai/unsloth/pull/2793
• intel 5/N fix patch by @mmathew23 in https://github.com/unslothai/unsloth/pull/2792
• [3/N] Enable intel GPU for unsloth by @leizhenyuan in https://github.com/unslothai/unsloth/pull/2620
• [4/N] Enable intel GPU for unsloth by @mmathew23 in https://github.com/unslothai/unsloth/pull/2801
• [intel] use DeviceProperties instead of torch.xxx.deviceproperties by @leizhenyuan in https://github.com/unslothai/unsloth/pull/2803
• Fix grpo sleep regex and indentation by @Datta0 in https://github.com/unslothai/unsloth/pull/2804
• Bug fixes by @danielhanchen in https://github.com/unslothai/unsloth/pull/2805
• Bug fixes by @danielhanchen in https://github.com/unslothai/unsloth/pull/2807

New Contributors

• @cblomert made their first contribution in https://github.com/unslothai/unsloth/pull/2452
• @johnnynunez made their first contribution in https://github.com/unslothai/unsloth/pull/2458
• @rolandtannous made their first contribution in https://github.com/unslothai/unsloth/pull/2416
• @yuanzhedong made their first contribution in https://github.com/unslothai/unsloth/pull/2492
• @mmathew23 made their first contribution in https://github.com/unslothai/unsloth/pull/2466
• @leizhenyuan made their first contribution in https://github.com/unslothai/unsloth/pull/2350
• @fenglui made their first contribution in https://github.com/unslothai/unsloth/pull/2158
• @kiankyars made their first contribution in https://github.com/unslothai/unsloth/pull/2537
• @omahs made their first contribution in https://github.com/unslothai/unsloth/pull/2540
• @jkumz made their first contribution in https://github.com/unslothai/unsloth/pull/2539
• @davedgd made their first contribution in https://github.com/unslothai/unsloth/pull/2563
• @emmanuel-ferdman made their first contribution in https://github.com/unslothai/unsloth/pull/2564
• @qgallouedec made their first contribution in https://github.com/unslothai/unsloth/pull/2609
• @Premik made their first contribution in https://github.com/unslothai/unsloth/pull/2633
• @RunFMe made their first contribution in https://github.com/unslothai/unsloth/pull/2216
• @Salpingopharyngeus made their first contribution in https://github.com/unslothai/unsloth/pull/2680
• @user799595 made their first contribution in https://github.com/unslothai/unsloth/pull/2637
• @pluesclues made their first contribution in https://github.com/unslothai/unsloth/pull/2710
• @leopardracer made their first contribution in https://github.com/unslothai/unsloth/pull/2721
• @marcandrelarochelle made their first contribution in https://github.com/unslothai/unsloth/pull/2774
• @amrothemich made their first contribution in https://github.com/unslothai/unsloth/pull/2778
• @simpissa made their first contribution in https://github.com/unslothai/unsloth/pull/2702
• @kilavvy made their first contribution in https://github.com/unslothai/unsloth/pull/2784

Full Changelog: https://github.com/unslothai/unsloth/compare/May-2025...June-2025