unsloth Releases

Minimal release notes: "## Changelog * d607566 chore: Update version to v1.4.1 in code and documentation"

Aug 24, 2025

Adds full support for fine-tuning gpt-oss models with significant VRAM reduction and speed improvements, alongside support for Qwen3 and new model variants.

Aug 8, 2025

Major VRAM reduction improvements and bug fixes across various models including Gemma 3N, Gemma 3, Falcon H1, and GRPO.

Jul 10, 2025

Introduces Gemma 3n (multimodal) and Text-to-Speech (TTS) fine-tuning support, along with DeepSeek-R1-0528 support.

Jun 26, 2025

Adds comprehensive support for Qwen3 models, including the 30B MoE variant, and enables full-finetuning and 8bit finetuning.

May 2, 2025

Introduces Gemma 3 support across various sizes and adds preliminary support for full-finetuning, 8bit finetuning, and Mixtral.

Mar 14, 2025

Enhances rendering capabilities by supporting cameras with non-centered principal points, simplifies the FlameViewer rendering pipeline, and adds temporal smoothness to expression terms.

Mar 7, 2025

Introduces GRPO (Group Relative Policy Optimization) with significant memory efficiency improvements and long context support.

Feb 20, 2025

Adds GRPO (Group Relative Policy Optimization) support and vLLM integration for fast inference. Includes Phi-4 and Llama 3.1 support with GRPO examples. Fixes bugs related to Gemma 2 and Mistral base mapping.

Feb 6, 2025

Introduces Phi-4 support and fixes critical bugs caused by new transformers versions. Includes llama-fied Phi-4 and dynamic 4bit quants.

Jan 10, 2025

Introduces dynamic 4bit quants for improved accuracy and adds support for Llama 3.3. Highlights extreme context length support.

Dec 4, 2024

Major update adding Vision finetuning capabilities for Llama 3.2 Vision, Pixtral, and Qwen2-VL.

Nov 21, 2024

Migrates the rendering backend from OpenGL to CUDA to utilize new features in nvdiffrast, enabling higher resolution outputs and simplifying the installation process.

Oct 29, 2024

Fixes a critical gradient accumulation bug that caused training losses to not match theoretical expectations.

Oct 15, 2024

Adds Qwen 2.5 support and fixes critical EOS token issues for base models.

Sep 23, 2024

Adds Phi 3.5 support and various bug fixes.

Aug 21, 2024

Adds support for Llama 3.1, claiming 2.1x faster training and 60% less VRAM usage.

Jul 23, 2024

Adds Mistral NeMo support, Ollama export, and CSV support. Introduces RoPE scaling.

Jul 19, 2024

Adds Gemma 2 support with 2x speed and 63% less VRAM. Adds continued pretraining support.

Jul 3, 2024

This release focuses on enhancing vision capabilities, introducing the Qdrant vector store, and improving overall performance through faster authentication and JSON mode improvements.

Jun 4, 2024