MLX v0.32.4 adds 1-bit quantization, stft, and Python 3.15 support

MLX v0.32.4, released on October 14, 2026, adds 1-bit affine quantization support on Metal, new FFT operations, and Python 3.15 compatibility. The release also includes a large batch of kernel optimizations and bug fixes across CPU, Metal, and CUDA backends.

Quantization and kernel updates

The new 1-bit affine quantization for Metal targets memory-constrained inference. For quantized matmul, the release tunes M1 dispatch for six-row large shapes, uses the fast qmv kernel for outputs not divisible by 8, and keeps split-K partials in float32 to preserve precision. A new MLX_CONV_WINOGRAD switch lets users disable Winograd for conv2d, and thin N and thin M GEMM kernels expand dispatch coverage for narrow matrices. CUDA gains fp_gather_qmv and gather_qmm_rhs_sm80 kernels, plus a reduced floor for fp_gather_qmv.

Correctness and stability fixes

Several fixes target view and slicing behavior, including negative-stride views, misaligned offsets, and contiguity flags. The release corrects vmap issues with scatter batch axes, inverse real FFT for odd lengths, and complex VJPs for inverse trig, cos, sqrt, and rsqrt. It also fixes float16 overflow in clip_grad_norm, empty unsigned sum and prod on Metal, and CPU float16 SIMD comparison masks. Memory-related fixes include foreign buffer ownership, pickle bfloat16 strides, and a cross-process race in the CPU compile cache. A Metal fence deadlock and a CUDA eval cleanup deadlock are resolved.

New features and API additions

MLX now supports stft and istft in mlx.core.fft, a fixed-size unique operation, and a minmax method following NumPy and JAX APIs. Python arrays gain __pos__ support, and arctan2, radians, and degrees accept scalar inputs. The release adds Python 3.15 and 3.15t support, plus Metal extension build support. The Array API Tests suite is reintroduced, and the codebase moves from Black to ruff for formatting.

For inference engineers, the 1-bit quantization and GEMM dispatch changes can reduce memory footprint and improve throughput on Apple silicon, while the CUDA kernel additions benefit NVIDIA GPU users. The numerous view and slicing fixes address correctness issues that could surface in complex model code. The stft and istft additions expand signal processing capabilities for audio models. Python 3.15 support ensures compatibility with upcoming Python releases, and the Metal extension build support enables custom kernel development.