Google's JAX Releases v0.12.0: Adds MXFP4 Quantization and Proxy Backend, Removes Legacy Remat API

JAX, the open-source machine learning framework developed by Google, released version 0.12.0 on October 19, 2026. The update focuses on enhancing computational efficiency and hardware compatibility, particularly for memory optimization in large-scale model training.

Key features include support for 'mxfp4' format (a low-bit quantization scheme sharing one scale factor per block of 32 elements) in jax.nn.get_scaled_dot_general_config, with native execution on AMD gfx950 GPUs and a dequantize-and-dot fallback elsewhere. Additionally, JAX now registers a 'proxy' backend, allowing users to connect to IFRT Proxy servers like Pathways on Cloud via environment variables, simplifying distributed computing setup.

The release includes breaking changes. Most notably, jax.custom_remat has been removed; developers must now use jax.custom_vjp.defremat for equivalent functionality. Furthermore, JAX no longer supports x86-64 CPUs with only AVX instruction sets, raising the minimum requirement to Intel Haswell or newer (circa 2013 onwards, requiring AVX2 and FMA). NumPy dependency has also increased, now requiring NumPy 2.2 or higher.