Hey Product Hunt! 👋 We're excited to launch **ThousandML Training Factory**! 🚀 Our goal at ThousandML is simple: **make fine-tuning powerful open models more accessible and affordable without sacrificing the quality of the starting model.** Most fine-tuning workflows rely on full-precision or higher-precision base models, which can be expensive in terms of GPU memory and compute. We believe **efficient fine-tuning starts with high-quality quantization.** That's why we built Training Factory around our own **Coded Quantization (CQ)** technology — a calibration-free 4-bit quantization method designed to preserve model quality while significantly reducing memory requirements. Combined with QLoRA, Training Factory lets you: ✅ Fine-tune high-quality 4-bit quantized models with lightweight LoRA adapters. ✅ Reduce GPU memory requirements compared to full-precision fine-tuning. ✅ Fine-tune models like Llama 3.1 8B, Llama 3.3 70B, and Qwen using your own datasets. ✅ Serve your fine-tuned adapters with vLLM. ✅ Explore and customize the workflow with our open-source toolkit. 🔗 GitHub: https://github.com/thousandml/tr... We're still early, and community feedback is incredibly valuable to us. **We'd especially love to hear from you:** 💬 What's your biggest challenge when fine-tuning open LLMs today — GPU costs, model quality, or tooling complexity? And if you give Training Factory a try, we'd love to hear how it works for you, including any issues or features you'd like us to prioritize. Thanks for checking out our launch! 🙌

ThousandML

Fine-tune smarter with a high-quality 4-bit quantized base

Fine-tune smarter with a high-quality 4-bit quantized base