August 2025 Summaries

Unsloth introduces new advancements in vision/multimodal reinforcement learning (RL) with the integration of Gemma 3 and Qwen2.5-VL, significantly enhancing speed and efficiency by reducing VRAM usage by 90% and increasing context lengths by 10 times without loss of accuracy. The update incorporates the GSPO algorithm and allows training on a free Colab T4 GPU, with additional support for vLLM VLM integration and a new Standby feature that improves RL training speed by up to 10% and increases context lengths without additional memory usage. Unsloth also supports direct fine-tuning of gpt-oss models, resolving previous inference issues with collaborative efforts from Hugging Face and OpenAI, while aligning training loss behavior across different GPU setups. The platform's enhancements make RL training faster and more memory-efficient, with benchmarks showing significant improvements in context lengths for models like Qwen3-32B and Llama-3.1-8B.

Unsloth has introduced Flex Attention for OpenAI's gpt-oss training, enabling significantly longer context lengths, reduced VRAM usage, and faster training speeds compared to other implementations, including Flash Attention 3. This advancement allows for training with up to 60K context length on 80GB VRAM using BF16 LoRA, and even longer with QLoRA, while addressing previous issues such as infinite training losses on float16 GPUs. Flex Attention provides customizable attention mechanisms, allowing for efficient training by utilizing attention sinks and sliding window techniques. The new capability to export QLoRA fine-tuned gpt-oss models to platforms like llama.cpp, vLLM, and Hugging Face, along with improved direct fine-tuning support, enhances the flexibility and usability of the models. Additionally, bug fixes have been implemented to ensure consistent training loss behavior across different GPU configurations.

DeepSeek-V3.1 is an updated hybrid reasoning model by DeepSeek, designed to rival major AI models like OpenAI's GPT-4.5 and Google's Gemini 2.5 Pro, offering significant size reduction from 720GB to 170GB through selective quantization. This model, which requires 715GB for its full 671B parameter version, can be run effectively using Unsloth's 1-bit Dynamic 2.0 GGUFs on popular inference frameworks like llama.cpp and Ollama. Recommended settings for optimal performance include setting the temperature to 0.6 to minimize repetition, using a top_p of 0.95, and a context length of 128K. The update also addresses chat template issues to improve compatibility with various engines and provides detailed instructions for running the model locally or via platforms like Hugging Face. Users are encouraged to leverage community resources and platforms for support and updates.

OpenAI's gpt-oss models have achieved state-of-the-art performance in text, reasoning, math, and code, with gpt-oss-120b and gpt-oss-20b models outperforming their predecessors while being more VRAM efficient, enabling their use on free platforms like Colab. The Unsloth team has introduced enhancements to the models, including faster fine-tuning and reduced VRAM requirements, and has uploaded various versions on Hugging Face. The Harmony library has been released for improved parsing and tokenization, offering an alternative to existing chat templates. Additionally, the team has been working on making the gpt-oss more efficient by implementing custom training functions and addressing issues with the MXFP4 format, which currently lacks support for training. Unsloth's optimizations allow for training the models with significantly lower VRAM usage compared to other methods, and collaboration with the Falcon and LiquidAI teams has expanded support to their latest models.