AI 导读
mlx-vlm v0.6.3 上线，当天即支持 Google DeepMind 的 DiffusionGemma 与 Cohere North Mini Code 1.0 在 Mac 本地 MLX 运行。DiffusionGemma 采用新架构，以 256 token 整块并行生成，配合双向注意力与迭代自纠错，26B MoE 仅激活 3.8B，量化后 18GB 可跑；North Mini Code 为 30B MoE、3B active，BF16 下约 66 tok/s。用户可通过 uv pip install -U mlx-vlm 安装，并以 uv run mlx_vlm.server --model MODEL-REPO 启动服务体验。
正文
Prince Canuma直接把Google刚发布的DiffusionGemma和Cohere North Mini Code当天塞进Mac本地MLX，零等待直接把玩咯！
mlx-vlm v0.6.3刚上线，DiffusionGemma这个新架构直接生成256 token整块、双向注意力+迭代自纠错，26B MoE只激活3.8B，量化后18GB就能跑。
North Mini Code 30B MoE也只要3B active，BF16下66 tok/s起步。
全靠和Google DeepMind、Cohere的深度合作，Day-0支持拉满！
一键安装即可体验啊~
地址：huggingface.co/collections/m…
mlx-vlm v0.6.3 is here 🚀 Day-0 support for TWO new models from our partners we work closely with: 🔥 @GoogleDeepMind DiffusionGemma — a genuinely new architecture. Instead of token-by-token, it generates 256-token blocks in parallel with bi-directional attention and iteratively self-corrects the whole block, image-generator style. 26B MoE, only 3.8B active, fits in 18GB quantized. Day-0 MLX support via our Google DeepMind partnership, with long-context prefill tuned and ready. 🔥 @cohere's North Mini Code 1.0 — a 30B MoE with just 3B active, running ~66 tok/s in BF16 before any compression. Day-0 on MLX thanks to our close collaboration with the Cohere team. Get started today — install from source: > uv pip install -U mlx-vlm Then serve the model and point your favorite agent at it (pi, opencode, hermes, etc.): uv run mlx_vlm.server --model MODEL-REPO Model collection 👇🏽在 X 查看被引用的帖子
来源：@berryxia · x.com