Post

Jumped from 20 TPS to 80 TPS ⚡ on my M4 Pro. The trick is that this version uses a MoE architecture with 3 billion active parameters. That makes it way faster. I was running the base model via GGUF (ollama), and now running on MLX (vlm - because this is a vision model too).

Sam Edwards

@handstandsam.bsky.social

2 Mar

Is that possible with ollama? I'm still new to local models and ollama makes that easy for me still. I saw there is Qwen3.5 ollama.com/library/qwen... and tried the lowest one on a M4 but was still slow out of the box.

ollama.comqwen3.5Qwen 3.5 is a family of open-source multimodal models that delivers exceptional utility and performance.

iury souza

@iurysouza.dev

That's what I tried before. It's 4x slower than the alternative.

2/3/26, 3:18 pm0 reposts1 like

Sam Edwards

@handstandsam.bsky.social

2 Mar

Gotcha, I think you can configure models and params with ollama, think that's possible here, or would need to go the more advanced llama method?

iury souza

@iurysouza.dev

2 Mar

Yeah, so the thing is the model is completely different. That one is the dense version of the 3.5:35b. I shared more details on linkedin. (should fit better than plain text here) 👇 www.linkedin.com/feed/update/...