Klose Li

[Klose Li @lwl53]

@Kurnalsalts 请问有M5系列 chiplet计划么？

[Klose Li @lwl53]

@JackdeS11 thank you for your feedback. i am freshman on swift. With the latest github code and hf model, qwen3.5-0.8b is working on my M4MAX ANE, and the decode speed is about 61 tokens/s. Appreciate a lot for your work.

[Klose Li @lwl53]

@lmstudiodevs long time since last mlx updated

[Klose Li @lwl53]

@ivanfioravanti What do you think of mxfp8 vs. int8? Most quantization relies on int8, but ChatGPT says that mxfp8 has better precision.

[Klose Li @lwl53]

@SkylerMiao7 Do you have plans to release a model under 100B for local deployment and excels at coding tasks?

[Klose Li @lwl53]

@0xSero thank you for your optimization. I have convert the reap-30 into mlx-mxfp4, it works great especially in tool calling.

[Klose Li @lwl53]

@auto_localize @lmstudio @MistralAI The vision API of LM Studio / mlx expects "image_base64":"AAAA..." , but cline provides the official OpenAI format shown as below
"image_url":{ "url":"data:image/png;base64,..."}

So LM Studio will not parse it.

[Klose Li @lwl53]

@auto_localize @lmstudio @MistralAI I have the same issue when Cline is using the browser tool

[Klose Li @lwl53]

@yagilb MLX 0.31.0 , mlx-vlm==0.3.4
When will be updated to latest 0.3.5? I would like to try Deepseek-OCR

[Klose Li @lwl53]

@UnicornBitcoin @HuoQubot

[Klose Li @lwl53]

It’s killing me to delete more than 300 tweets one by one, fxcking the guy hacked my Twitter

[Klose Li @lwl53]

国内能上twitter youtube和facebook了！mark一下