Skip to content

Can we add DeepSeek V4 to mlx_lm #1281

Description

@rajveer43

DeepSeek-V4-Flash (284B total / 13B activated) represents the most efficient large MoE model released to date — achieving 9.8× lower FLOPs and 13.7× smaller KV cache than DeepSeek-V3.2 at 1M-token context. Its activated parameter count (13B) makes it practically runnable on Apple Silicon with quantization, yet it punches well above its weight class on reasoning, code, and long-context benchmarks.

mlx-lm currently supports DeepSeek-V2 and V3 architectures. V4 introduces two new attention mechanisms — Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) — that are not yet implemented. Without native mlx-lm support, Apple Silicon users cannot run the most capable open-source MoE model available.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions