Daniel Han is the co-founder of Unsloth, the open-source AI company he founded with his brother Michael to make powerful models easier to fine-tune, compress, and run locally. His work addresses a practical constraint on open AI: useful models often exceed ordinary developers’ hardware budgets, while subtle implementation errors can quietly undermine their accuracy.
From GPU acceleration to Unsloth
Han’s earlier projects included sciblox, a toolkit for easier data science and machine learning, and an investigation into reversing Markov chains. He and Michael also developed HyperLearn, an earlier machine-learning software project.
At NVIDIA, Han worked on GPU-accelerated machine learning in the RAPIDS ecosystem. His article about accelerating t-SNE described improvements to the dimensionality-reduction algorithm in RAPIDS cuML, including substantial speedups under specific testing conditions.
The brothers launched Unsloth in December 2023 to accelerate language-model fine-tuning and reduce memory consumption without requiring new hardware. Unsloth joined Y Combinator’s Summer 2024 batch and expanded into reinforcement learning, model distribution, quantization, and local AI workflows.
- Open-model debugging: Han identified failures in Gemma, Llama, and Phi models that stemmed from activation-function discrepancies, incorrect numerical precision, duplicate beginning-of-sequence tokens, incompatible chat templates, untrained token embeddings, and collisions between padding and end-of-sequence tokens. Such defects can cause invalid gradients, degraded fine-tuning, or endless generation. His technical walkthrough of open-model failures treats tokenization, training, inference, and export as one interconnected system.
- Correct gradient accumulation: Daniel and Michael traced discrepancies between accumulated small-batch training and equivalent full-batch training to incorrect loss normalization across sequences of different lengths. Their gradient-accumulation investigation addressed a mathematical error embedded in widely used training workflows and supported related improvements with Hugging Face.
- Asynchronous gradient checkpointing: The brothers’ long-context training research describes moving selected intermediate activations into system memory without unnecessarily blocking GPU execution. This approach makes longer fine-tuning contexts feasible while limiting additional runtime.
- Dynamic quantization: Han’s compression strategy preserves sensitive layers or weights at higher precision while aggressively compressing less sensitive components. Unsloth’s dynamic four-bit research demonstrates why uniform compression can damage a vision-language model’s interpretation even when the resulting file is smaller. In an August 2026 update, Han reported one-bit Qwen variants capable of running in 8 GB of RAM, with retained accuracy measured through Unsloth’s internal testing.
- Reward design and benchmark reliability: Han emphasizes that reinforcement learning depends on meaningful reward functions, adequate exploration, and verification that reflects the intended task. His reasoning-model workshop connects supervised fine-tuning, GRPO, parameter-efficient adaptation, and low-memory inference. His work on reward hacking and evaluation examines how weak tests, contaminated benchmarks, misleading timing measurements, and incorrect intermediate reasoning can produce impressive scores without corresponding improvements.
Han’s goal is democratizing capable local AI through software that preserves useful behavior while reducing the memory, hardware, and expertise required to adapt advanced open models.
▶ Watch ↗
