Hacker News Reader

Comment by omneity | Hacker News Reader

I'm using Qwen3-30B-A3B locally and it's very impressive. Feels like the GPT-4 killer we were waiting for for two years. I'm getting 70 tok/s on an M3 Max, which is pushing it into the "very usable" quadrant.

What was even more impressive is the 0.6B model which made the sub 1B actually useful for non-trivial tasks.

Overall very impressed. I am evaluating how it can integrate with my current setup and will probably report somewhere about that.

View on HN

Personally, I'm getting 15 tok/s on both RTX 3060 and my MacBook Air M4 (w/ 32GB, but 24 should suffice), with the default config from LMStudio.

Which I find even more impressive, considering the 3060 is the most used GPU (on Steam) and that M4 Air and future SoCs are/will be commonplace too.

(Q4_K_M with filesize=18GB)

One of the most interesting things about that model is its excellent score on the RAG confabulations (hallucination) leaderboard. It’s the 3rd best model overall, beating all OpenAI models, for example. I wonder what Alibaba did to achieve that.

https://github.com/lechmazur/confabulations

What tasks have you found the 0.6B model useful for? The hallucination that's apparent during its thinking process put up a big red flag for me.

Conversely, the 4B model actually seemed to work really well and gave results comparable to Gemini 2.0 Flash (at least in my simple tests).

You can use 0.6B for speculative decoding on the larger models. It'll speed up 32B, but slows down 30B-A3B dramatically.

It's okay for extracting simple things like addresses, or for formatting text with some input data, like a more advanced form of mail merge.

I haven't evaled these tasks so YMMV. I'm exploring other possibilities as well. I suspect it might be decent at autocomplete, and it's small enough one could consider finetuning it on a codebase.

Importantly they note that using a draft model screws it up and this was my experience. I was initially impressed, then started seeing problems, but after disabling my draft model it started working much better. Very cool stuff- it's fast too as you note.

The /think and /no_think commands are very convenient.

What do you mean by draft model? And how would one disable it? Cheers

A draft model is something that you would explicitly enable. It uses a smaller model to speculatively generate next tokens, in theory speeding up generation.

Here’s the LM Studio docs on it: https://lmstudio.ai/docs/app/advanced/speculative-decoding

That should not be the case. Speculative decoding is trading off compute for memory bandwidth. The model's output is guaranteed to be the same, with or without it. Perhaps there's a bug in the implementation that you're using.

how much RAM do you have? I want to compare with my local setup (M4 Pro)

128GB but it's not using much.

I'm running Q4 and it's taking 17.94 GB VRAM with 4k context window, 20GB with 32k tokens.

I am not a mac person, but I am debating buying one for the unified ram now that the prices seem to be inching down. Is it painful to set up? The general responses I seem to get range from "It is takes zero effort" to "It was a major hassle to set everything up."

Read the article you are commenting on. It is a how to that answers your exact question. It takes 4 commands in the terminal.

The article should answer your question. Or do you mean setting up a Mac for use as a Linux or windows user

<< Or do you mean setting up a Mac for use as a Linux or windows user

This part, yes. I assume the setting a complete environment is a little more involved than the 4 commands sibling is also refers to.

I mean that's more like old habits no? MacOS is pretty easy to setup. if all your target apps are available on it, you shouldn't have much of a problem.

As a Windows/MacOS/Linux dweller kinto is a godsend so I can have macos keyboard (but you could have linux or windows by default) on all OSes https://kinto.sh/

It's relatively easy. macOS is easier to set up than Linux, but it will always depend on your specific needs and environment.

E.g.: I go a little bit overboard for the average macOS user:

- custom system- and app-specific keyboard mappings (ultra-modifier on caps-lock; custom tabbing-key-modifier) via Karabiner Elements

- custom trackpad mappings via BetterTouchTool

- custom Time Machine schedule and backup logic; you can vibe-code your install script once and re-use it in the future; just make it idempotent

- custom quake-like Terminal via iTerm

- shell customizations

- custom Alfred workflows

- etc.

If all you need is just a sensible package manager and the terminal to get started, just set up Time Machine with default settings, Homebrew, your shell, and optionally iTerm2, and you're good to go. Other noteworthy power-user tools:

- Hammerspoon

- Syncthing / Resilio Sync

- Arq. Naturally, the usual backup tools also run on macOS: Borg, Kopia, etc.

- Affinity suite for image processing

- Keyshape for animations on web, mobile, etc.

LM Studio and Ollama are both very low complexity ways to get local LLMs running on a Mac.

As a Python person I've found uv + MLX to be pretty painless on a Mac too.

You can use the method in this tutorial or you can download LM Studio and run it.

The latter is super easy. Just download the model (thru the GUI) and go.

Honestly, it is quite a hastle, took me 2 hours BUT. if you just take the whole article text and paste that to gemini-2.5-pro and give your circumstance, i think it will give you specific steps for your case and it should be trivial from that moment on

> I am not a mac person, but I am debating buying one for the unified ram

Soon some AMD Ryzen AI Max PCs will be available, with unified memory as well. For example the Framework Desktop with up to 128 GB, shared with the iGPU:

- Product: https://frame.work/us/en/desktop?tab=overview

- Video, discussing 70B LLMs at around 3m:50s : https://youtu.be/zI6ZQls54Ms

oh boy... i was genuinely hoping someone less pricy would enter this market

edit: ok.. i am excited.

Using llama.cpp or pytorch could hardly be easier.

I have a MBP M1 Max 64GB and I get 40t/s with llama.cpp and unsloth q4_k_m on the 30B A3B model. I always use /nothink and Temperature=0.7, TopP=0.8, TopK=20, and MinP=0 - these are the settings recommended for Qwen3 and they make a big difference. With the default settings from llama-server it will always run into an endless loop.

The quality of the output is decent, just keep in mind it is only a 30B model. It also translates really well from french to german and vice versa, much better than Google translate.

Edit: for comparision, Qwen2.5-coder 32B q4 is around 12-14t/s on this M1 which is too slow for me. I usually used the Qwen2.5-coder 17B at around 30t/s for simple tasks. Qwen3 30B is imho better and faster.

[1] parameters for Qwen3: https://huggingface.co/Qwen/Qwen3-30B-A3B

[2] unsloth quant: https://huggingface.co/unsloth/Qwen3-30B-A3B-GGUF

[3] llama.cpp: https://github.com/ggml-org/llama.cpp

I am curious if you have tried function calling or MCPs with it?

It's fits entirely in my 7900xtx memory. But tbh i've been disappointed with programming ability so far.

It's using 20GB of memory according to ollama.

I’m getting 56 with mlx and lmstudio. How 76?

I went form bottom up, started with 4B, then 8B, then 30B, and 30B was the only one that started to "use tools". Other models were saying they will use but never did, or didnt notice all the tools. I think anything above 30b would atually be able to go full GPT on a task. 30b does it, but a bit.. meh

Reply on news.ycombinator.com