A new MIT-licensed tool, lora-merge, was released on GitHub on 2026-10-18, offering a GUI for merging up to five LoRA adapters into a base model with per-adapter weight sliders. It ships as a portable 100 MB offline Windows executable and exports safetensors.
A new arXiv paper examines why merging a LoRA adapter into a native 4-bit microscaling checkpoint (NVFP4, MXFP4) is no longer a free deployment step. Merged weights must be re-quantized, coupling the deployed artifact to one quantization convention. The paper proposes a code-invariant merging approach for these formats.
A new arXiv paper examines why merging a LoRA adapter into a native 4-bit microscaling checkpoint (NVFP4, MXFP4) is no longer a free deployment step. Merged weights must be re-quantized, coupling the deployed artifact to one quantization convention. The paper proposes a code-invariant merging approach for these formats.
A developer created a post-training delta compression tool that stores only the weight differences between a fine-tuned model and its base, reducing storage for 10 fine-tunes to roughly the size of 4 full checkpoints. The tool works on any trained model without requiring changes during training, unlike LoRA.
A developer created a post-training delta compression tool that stores only the weight differences between a fine-tuned model and its base, reducing storage for 10 fine-tunes to roughly the size of 4 full checkpoints. The tool works on any trained model without requiring changes during training, unlike LoRA.
A developer released FTRAIN v1.0, an open-source AI framework built on Unsloth that aims to reduce VRAM usage during fine-tuning and prevent model merge crashes. It features an LLM-based 'Captain' supervision layer for intelligent training.
A developer released FTRAIN v1.0, an open-source AI framework built on Unsloth that aims to reduce VRAM usage during fine-tuning and prevent model merge crashes. It features an LLM-based 'Captain' supervision layer for intelligent training.
A developer known as UnstableLlama released exl3-qlora, a fork of ExLlamaV3 implementing native QLoRA-style training on 2-8bpw EXL3 quantized models in plain PyTorch. It breaks the previous BitsAndBytes 4-bit lock-in, supports SFT/DPO/KTO across major model families, and reports lower eval loss than bnb NF4 at comparable VRAM. The tool is an unofficial community project already used to produce released models.
A developer known as UnstableLlama released exl3-qlora, a fork of ExLlamaV3 implementing native QLoRA-style training on 2-8bpw EXL3 quantized models in plain PyTorch. It breaks the previous BitsAndBytes 4-bit lock-in, supports SFT/DPO/KTO across major model families, and reports lower eval loss than bnb NF4 at comparable VRAM. The tool is an unofficial community project already used to produce released models.
Unsloth now officially supports AMD hardware for local inference, fine-tuning, reinforcement learning, and deployment. The open-source tool works on Windows, Linux, and WSL with Radeon RX 9000/7000 series, Instinct MI350/MI300 GPUs, Strix Halo/Ryzen AI Max systems, and AMD CPUs for GPU-free inference. It claims up to 70% less VRAM for training and 80% less for reinforcement learning.
Unsloth now officially supports AMD hardware for local inference, fine-tuning, reinforcement learning, and deployment. The open-source tool works on Windows, Linux, and WSL with Radeon RX 9000/7000 series, Instinct MI350/MI300 GPUs, Strix Halo/Ryzen AI Max systems, and AMD CPUs for GPU-free inference. It claims up to 70% less VRAM for training and 80% less for reinforcement learning.