New version 0.5.10

You have 0.5.5. Found when this page was made.

What's new:

Write a component as a :::callout … ::: block too; it renders like ```

Run in a terminal:

npx skills update answer-me-with-html -y

Or tell your agent:

Update answer-me-with-html

Can uzu speed up your open models?

On this Mac today: it won't run

uzu's GPU code uses Metal 4 (MTL4CommandQueue), and its build targets macOS 26.4. This Mac runs macOS 14.6 with Xcode 16.2, so neither the package nor a build from source will work. The M1 Pro itself is supported by macOS 26, so upgrading the OS (and Xcode 26) would unblock it.

Estimate after an upgrade (Qwen 3.5 9B, 4-bit)

Engine	Expected tok/s on M1 Pro	Basis
llama.cpp / MLX	~25–33	200 GB/s ÷ 5.2 GB ≈ 38 ceiling, engines reach ~70–85%
uzu (speculative)	~50–80, unproven	~7.5 tokens per pass, but the M1 has no int8 matrix path, so each verify pass costs more

Expect roughly 2–3x, not the 4x from the post. Nobody has published numbers for M1.

Why the M1 gets less

Choices

Upgrade to macOS 26 + Xcode 26, then I run uzu bench against llama.cpp on Qwen 3.5 9B for a real number.
Stay on macOS 14: llama.cpp also has speculative decoding (-md draft model), usually ~1.5–2x on Qwen, and it works today.

---
title: Can uzu speed up your open models?
---

# On this Mac today: it won't run

uzu's GPU code uses Metal 4 (`MTL4CommandQueue`), and its build targets macOS 26.4. This Mac runs **macOS 14.6** with **Xcode 16.2**, so neither the package nor a build from source will work. The M1 Pro itself is supported by macOS 26, so upgrading the OS (and Xcode 26) would unblock it.

# Estimate after an upgrade (Qwen 3.5 9B, 4-bit)

| Engine | Expected tok/s on M1 Pro | Basis |
|---|---|---|
| llama.cpp / MLX | ~25–33 | 200 GB/s ÷ 5.2 GB ≈ 38 ceiling, engines reach ~70–85% |
| uzu (speculative) | ~50–80, unproven | ~7.5 tokens per pass, but the M1 has no int8 matrix path, so each verify pass costs more |

Expect roughly **2–3x**, not the 4x from the post. Nobody has published numbers for M1.

# Why the M1 gets less

```flow
Drafter guesses 16 tokens -> 9B checks the whole tree in one pass -> on M5: int8 path keeps the check cheap -> on M1: fp16 compute, check costs ~1.5-2x a normal pass
```

# Choices

- **Upgrade to macOS 26 + Xcode 26**, then I run uzu `bench` against llama.cpp on Qwen 3.5 9B for a real number.
- **Stay on macOS 14**: llama.cpp also has speculative decoding (`-md` draft model), usually ~1.5–2x on Qwen, and it works today.