Skip to main content
Back

Intel's AutoRound v0.14 Shrinks 200B+ Models Without a Single GPU

Intel's AutoRound quantization toolkit gets model-free quantization, stronger MXFP4, diffusion model support, and multi-GPU scaling in v0.14

Repo, Full Access
  • v0.14 announced: Intel's AutoRound quantization toolkit adds model-free quantization, improved MXFP4, diffusion model support, and multi-GPU quantization.
  • Model-free mode: auto-round-rtn now defaults to calibration-free quantization — no GPU required to compress a model.
  • MXFP4 improvements: Hadamard transform support added for MXFP4, improving accuracy for this OCP-standard 4-bit floating-point format.
  • Wider model coverage: WAN2.2 diffusion models, Qwen-TTS, MiMo-Audio, BAGEL-7B-MoT, Gemma4, and more are now supported.
  • Ecosystem integration: AutoRound is integrated into vLLM, SGLang, Transformers, and LLM Compressor for direct deployment of quantized models.
  • Hardware roadmap alignment: Intel's upcoming Crescent Island GPU adds native MXFP4/MXFP8 support, making AutoRound-quantized models forward-compatible.

Intel's AutoRound , a post-training quantization toolkit for LLMs and vision-language models , has a new release on the way. Intel researcher HaihaoShen announced AutoRound v0.14 on X, highlighting four headline additions: model-free quantization, improved MXFP4, WAN/diffusion model support, and multi-card quantization. The release follows a rapid cadence of updates that has seen the project grow to over 1,500 GitHub stars.

AutoRound's core idea is to make quantization smarter than simple rounding. It introduces three trainable parameters per quantized tensor , a rounding offset, and learned clipping range controls , then applies signed gradient descent to jointly optimize rounding and clipping, minimizing block-wise output reconstruction error. The result is that models compressed to 2–4 bits retain far more accuracy than naive approaches.

What's new in v0.14

The four features called out in the release announcement each address a real friction point in production quantization workflows:

  • Model-free quantization. AutoRound now provides free devices for calibration-free quantization via pure RTN mode, and auto-round-rtn