Post

Post

• 

  🍎⚡ Qwen3.8-27B is now hitting ~70 tok/s on an M5 Max MacBook Pro with MLX. And this isn't a smaller model or a quality-reducing speed hack. DFlash 2 landed for Qwen3.8-27B, using speculative decoding to generate multiple candidate tokens in parallel while Qwen verifies them. This results in much faster local generation while preserving the target model's output. Apple Localmaxxers can run it now with oMLX. You need TWO models: 🧠 Target model mlx-community/Qwen3.8-27B-4bit ⚡ DFlash 2 drafter incoai/Qwen3.8-27B-DFlash2 How to run it 1️⃣ Install the new oMLX build with DFlash 2 support 2️⃣ Open oMLX Model Downloader Download: mlx-community/Qwen3.8-27B-4bit and: incoai/Qwen3.8-27B-DFlash2 3️⃣ In oMLX Model Manager, edit Qwen3.8-27B and set: ✅ DFlash: Enabled ✅ Draft model: incoai/Qwen3.8-27B-DFlash2 ✅ Draft quantization: Enabled ✅ Runtime block size: 5 ✅ Verify mode: dflash Save → load Qwen3.8 → run. That's it. A 27B Qwen running locally on a MacBook at around 70 output tok/s. No API. No NVIDIA GPU 🙂. No cloud inference. ✊ Just Apple unified memory + MLX + a much smarter way of decoding.

  

• 

  I ran a DFlash 2 matrix yesterday and found one caveat: long outputs did not stay byte-identical to plain autoregressive decoding. Prose diverged after 554 characters and code after 974, while arithmetic, tool calls and vision stayed identical across three repeats. This post

  

  I spent a day trying to squeeze more juice out of a Mac mini M4 Pro (64 GB) with Qwen3.8-27B. The question was simple: on this chip, what actually makes the model faster, and what only looks fast? A few words first, then the numbers. A token is a piece of a word. tok/s is how

• 

  If we can only get prefill throughput up

• 

  Anyone tried on an M1?