mlx-lm 0.32 adds Muse Spark Open NVFP4; MLX 0.32.4 trims pre-M5 SDPA D512 memory

ml-explore shipped three mlx-lm releases and one MLX release between October 14 and 16, 2026. mlx-lm 0.32.0 (Oct 15) adds `meta-models/Muse-Spark-Open-816B-A42B-NVFP4` (#1976). Going by the name, the model has 816B total and 42B active parameters. The release notes do not state parameter counts or a memory requirement. The same release adds a global scale for NVFP4-quantized MoEs (#1844) and `prism_hadamard_qwen35`, a Hadamard-rotated 2-bit Qwen3.5 (#1941). It also adds support for unconverted models (#1962), fixes a long-generation crash with MiniMax M3 (#1958) and adds a tool-call parser for MiniCPM5 (#1899). mlx-lm 0.32.1 (Oct 15) has one change: it fixes the Muse template (#1977). mlx-lm 0.32.2 (Oct 16) fixes attention value scaling in MiMo-V2-Flash (#1971), fixes a mask in SMM (#1979) and updates the Muse Spark Hugging Face names (#1981). MLX 0.32.4 (Oct 14) improves Metal memory use for SDPA D512 on pre-M5 chips (#4518). The notes do not define D512; it most likely means an attention head dimension of 512. The release also tunes M1 quantized-matmul dispatch for six-row large shapes (#4629), adds Python 3.15 and 3.15t support (#4555) and fixes cooperative-tensor build issues on Mac OS 27 (#4594). For the CUDA backend it adds `fp_gather_qmv` to optimize `gather_qmm` (#4508) and adds `gather_qmm_rhs_sm80` (#4554).