Qwen3

Qwen 3 support + bug fixes

Please update Unsloth via pip install --upgrade --force-reinstall unsloth unsloth_zoo

Qwen3 notebook: https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/Qwen3_(14B)-Reasoning-Conversational.ipynb
GRPO with Qwen3 notebook: https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/Qwen3_(4B)-GRPO.ipynb

There are also many bug fixes in this release!

The 30B MoE is also fine-tunable in Unsloth!

from unsloth import FastModel
import torch
model, tokenizer = FastModel.from_pretrained(
    model_name = "unsloth/Qwen3-30B-A3B",
    max_seq_length = 2048, # Choose any for long context!
    load_in_4bit = True,  # 4 bit quantization to reduce memory
    load_in_8bit = False, # [NEW!] A bit more accurate, uses 2x memory
    full_finetuning = False, # [NEW!] We have full finetuning now!
    # token = "hf_...", # use one if using gated models
)

What's Changed

• GGUF saving by @danielhanchen in #2017
• Gemma 3 readme by @danielhanchen in #2019
• Update README.md by @danielhanchen in #2028
• bug fix #2008 - load_in_4bit = True + fast_inference = True by @void-mckenzie in #2039
• unsloth_fast_generate model is not defined fix by @KareemMusleh in #2051
• Ensure trust_remote_code propagates down to unsloth_compile_transformers by @CuppaXanax in #2075
• Show peft_error by @IsaacBreen in #2080
• Add generation prompt error message change by @KareemMusleh in #2046
• Many bug fixes by @danielhanchen in #2087
• fix: config.torch_dtype in LlamaModel_fast_forward_inference by @lurf21 in #2091
• Updating new FFT 8bit support by @shimmyshimmer in #2110
• Bug fixes by @danielhanchen in #2113
• Small fix by @danielhanchen in #2114
• fix(utils): add missing importlib import to fix NameError by @naliazheli in #2134
• Add QLoRA Train and Merge16bit Test by @jeromeku in #2130
• Fix Transformers 4.45 by @danielhanchen in #2151
• Bug Fixes by @danielhanchen in #2197
• Issues templates by @jeromeku in #2242
• Fix feature_request ISSUE_TEMPLATE by @jeromeku in #2250
• Registry refactor by @jeromeku in #2255
• Update README.md by @Kimizhao in #2267
• Update README.md by @jackswl in #2119
• Update bug_report.md by @shimmyshimmer in #2323
• feat: Support custom auto_model for wider model compatibility (Whisper, Bert,etc) & attn_implementation support by @Etherll in #2263
• fix: improved error handling when llama.cpp build fails by @Hansehart in #2358
• Revert "fix: improved error handling when llama.cpp build fails" by @shimmyshimmer in #2375
• Fix saving 4bit for VLM by @Erland366 in #2381
• [WIP] Initial support for Qwen3. Will udpate when the model is released by @Datta0 in #2211
• Fixup qwen3 by @Datta0 in #2423
• Fixup qwen3 qk norm by @Datta0 in #2427
• Qwen3 inference fixes by @Datta0 in #2436
• Update mapper.py to add Qwen3 base by @Etherll in #2439
• Qwen 3, Bug Fixes by @danielhanchen in #2445

New Contributors

• @void-mckenzie made their first contribution in #2039
• @CuppaXanax made their first contribution in #2075
• @IsaacBreen made their first contribution in #2080
• @lurf21 made their first contribution in #2091
• @naliazheli made their first contribution in #2134
• @jeromeku made their first contribution in #2130
• @Kimizhao made their first contribution in #2267
• @jackswl made their first contribution in #2119
• @Etherll made their first contribution in #2263
• @Hansehart made their first contribution in #2358

Full Changelog: 2025-03...May-2025