http://du.kokura.dehttp://du.kokura.de/posts/341http://du.kokura.de/posts/341Wed, 25 Mar 2026 04:51:15 GMT<div class="tgme_widget_message_forwarded_from accent_color">Forwarded from <a class="tgme_widget_message_forwarded_from_name" href="https://t.me/zhetengsha/5175" target="_blank" rel="noopener"><span>折腾啥</span></a> (<span class="tgme_widget_message_forwarded_from_author">小一</span>)</div><u>JANG: 面向 Apple Silicon 的 MLX 混合精度量化格式</u><br /><br />• 针对 MoE 与超大模型做按张量分配比特宽度，在接近 MLX 体积下优先保留 attention 与 router 精度，低比特场景下稳定性和效果更强<br /><br />• 支持 Qwen、Nemotron、MiniMax、DeepSeek 等架构，部分模型可在 16 GB 或 64 GB Mac 上运行，甚至实现 397B 级模型在 128 GB Mac 上推理<br /><br />• 原生兼容 MLX 生态，提供推理模式、VLM 支持、bfloat16 自动检测和开发者集成方案，适合在 Apple Silicon 上部署高压缩本地模型<br /><br /><a href="https://github.com/jjang-ai/jangq" target="_blank" rel="noopener" title="https://github.com/jjang-ai/jangq">https://github.com/jjang-ai/jangq</a><br /><br /><a href="/search/result?q=%23AppleSilicon" title="#AppleSilicon">#AppleSilicon</a> <a href="/search/result?q=%23MLX" title="#MLX">#MLX</a> <a href="/search/result?q=%23%E6%A8%A1%E5%9E%8B%E9%87%8F%E5%8C%96" title="#模型量化">#模型量化</a> <a href="/search/result?q=%23%E6%B7%B7%E5%90%88%E7%B2%BE%E5%BA%A6%E9%87%8F%E5%8C%96" title="#混合精度量化">#混合精度量化</a> <a href="/search/result?q=%23%E6%9C%AC%E5%9C%B0%E5%A4%A7%E6%A8%A1%E5%9E%8B" title="#本地大模型">#本地大模型</a> <a href="/search/result?q=%23MoE" title="#MoE">#MoE</a> <a href="/search/result?q=%23%E6%8E%A8%E7%90%86%E4%BC%98%E5%8C%96" title="#推理优化">#推理优化</a> <a href="/search/result?q=%23Mac" title="#Mac">#Mac</a> <a href="/search/result?q=%23AI" title="#AI">#AI</a> <a href="/search/result?q=%23GitHub" title="#GitHub">#GitHub</a><a class="tgme_widget_message_link_preview" href="https://github.com/jjang-ai/jangq" target="_blank" rel="noopener" title="JANG — GGUF for MLX. YOU MUST USE JANG_Q RUNTIME. Adaptive Mixed-Precision Quantization + Runtime for Apple Silicon - jjang-ai/jangq"> <div class="link_preview_site_name accent_color">GitHub</div> <img class="link_preview_image" alt="GitHub - jjang-ai/jangq: JANG — GGUF for MLX. YOU MUST USE JANG_Q RUNTIME. Adaptive Mixed-Precision Quantization + Runtime for…" src="/static/https://cdn4.telesco.pe/file/W5XUyYm_F7qJZtMeV2KreGLdWrJJPtIQPdmhIog_I2qoA1A9KfUNZajb-D-RL9ZDemj_YNG7Hq78kfiW0CJh8uU3tbpEfKGJdM0SLHJD1NyGLkfz6DCaFOVYbOSeP0-iHute5za49mJ5etnim-UtxeZdHewOxM2tiYOib76GMw2CTXu4BbLLcCrsO33trN2rOB_36qEjfx6yNleYOxGehF52WsosBl8NdP8Vq7lCeJ9qrRP58I1V0oLQeye9guI__YFS2qaroa89ZuX3160dL7QKlSyBy8Ipq1utx-QNqeUtw1VBwiac7qkPOhfaiXE9XLdDO468GJFm1U3R0LVYZw.jpg" width="1200" height="630" loading="eager" /> <div class="link_preview_title">GitHub - jjang-ai/jangq: JANG — GGUF for MLX. YOU MUST USE JANG_Q RUNTIME. Adaptive Mixed-Precision Quantization + Runtime for…</div> <div class="link_preview_description">JANG — GGUF for MLX. YOU MUST USE JANG_Q RUNTIME. Adaptive Mixed-Precision Quantization + Runtime for Apple Silicon - jjang-ai/jangq</div> </a>