Executable programs and extensions
Husky on Mac: Woof, MLX and the 730 tok/s result
A slow local reply can make a new Mac look tempting. Husky takes another route: tailoring the engine to one model, Woof. Before reading 730 tok/s as a hardware promise, look at the task that produced it.
Key takeaways
- 01
- Husky targets Woof on Apple silicon; it is not a universal LLM speed switch.
- 02
- 730 tok/s is the developer's M5 Max result for editing a function.
- 03
- Try it through Underdog for Mac; check Korean quality and your workload separately.
Husky, Woof and Underdog are different pieces
Underdog is the app, Woof supplies the answers, and Husky runs the model. The September 20 announcement compared identical 4-bit Woof 4B weights on an M5 Max with a 40-core GPU and 128GB memory.
Think of reading a book at home. The content resembles the model; finding, opening and turning its pages resembles the engine. Faster page turns do not improve the book's facts. Judge execution speed separately from answer quality.
The task behind 730 tok/s
The announcement lists function editing at 163 tok/s in MLX, 614 in Husky and 730 with Flash. A short email was 151, 164 and 260 respectively. Both improved, but by different amounts. The developer compared medians of three runs using identical weights and greedy decoding.
When the prompt already contains the code being edited, much of the reply repeats it. A new essay has less text to reuse. Ask how much of your output repeats the input. Applying the headline maximum to a long Korean explanation or another model erases that distinction.
Does Flash replace the model's answer?
Husky can propose tokens from earlier input or use a small draft trained for Woof. Woof verifies the proposals, while the engine chooses between lookup and drafting. This Flash is not another vendor's similarly named model.
It resembles correcting an existing draft rather than starting on blank paper. That can save work, but repeatedly wrong suggestions still cost time to check. A larger proposal block is not automatically faster. The same tradeoff helps explain speculative decoding.
First-token latency needs matching conditions
Conversation caching matters when continuing a chat. Reading a long document for the first time is different from reading only the latest sentence. Separate time to first token, subsequent generation speed and initial model loading.
For a practical comparison, record both a fresh conversation with the same document and a follow-up question. This preserves the benefit of fast continuation without implying the same wait for every new PDF.
How to try it on a Mac
The announced route is the Underdog Mac app. Check its current OS requirements and installation instructions, and grant mail or calendar access only as needed. Local inference does not mean every connected external service stops communicating.
Start with a short, non-sensitive text. Test typo correction, tone changes and action-item extraction separately against the original. Check omitted meaning and Korean politeness as well as speed before trusting it with your work.
What should Qwen or oMLX users change?
The Husky announcement is not a reason to replace a working Qwen server. We have not verified a public general-purpose setting that applies Woof's specialized engine to Qwen. If model choice and existing API integrations matter, keep the route that already provides them.
The buying lesson is modest: before replacing a slow-feeling machine, check whether the model and runtime suit the task. Our speed experience helps compare other hardware and model choices; it does not reproduce Husky's Woof measurements or multiply every model by 4.5.