# Simon Willison: Qwen3.6-35B-A3B on my laptop drew me a better pelican than Claude Opus 4.7

> A laptop-sized open-weight model drew a better pelican than Anthropic's frontier closed model — Simon Willison's classic SVG benchmark finally breaks the correlation.

- Org: Simon Willison
- Date: 2026-04-16
- Category: article
- Source: https://simonwillison.net/2026/Apr/16/qwen-beats-opus/
- Canonical page: https://ai-tldr.dev/releases/simonw-qwen-beats-opus/

## In short

Simon Willison runs the 21GB Qwen3.6-35B-A3B quant locally on a MacBook Pro M5 and finds it beats Anthropic's Opus 4.7 at the pelican-riding-a-bicycle SVG benchmark — and at flamingo-on-a-unicycle too.

## What it is

A blog post from Simon Willison, one of the AI community's most-read commentators, comparing Alibaba's new Qwen3.6-35B-A3B open-weight model with Anthropic's freshly-shipped Claude Opus 4.7. The comparison uses Willison's long-running 'pelican riding a bicycle' SVG benchmark — a quick-and-dirty visual test he has been running against every new model for over a year.

## How it works

Willison ran a 20.9 GB Q4_K_S quantization of Qwen3.6-35B-A3B (a 35B-parameter Mixture-of-Experts with 3B active) locally on a MacBook Pro M5 via LM Studio and the llm-lmstudio plugin. He hit Opus 4.7 through the Anthropic API, including its max-thinking mode. The Qwen SVG renders a recognizable bicycle and a pelican with a pouch; Opus 4.7 botches the bicycle frame in both standard and max-thinking attempts. A secondary 'flamingo on a unicycle' test gives the same result, which he argues rules out the 'they trained on the pelican benchmark' explanation.

## Why it matters

The pelican benchmark has historically tracked overall model quality, so a 21 GB local model beating a fresh frontier closed model is news for people tracking where the open-weight frontier actually is. It is also an argument against the benchmark as a serious eval — Willison explicitly flags this — while still being the kind of grounded, fetched-and-verified spot-check this community listens to.

## Who it's for

ML practitioners, local-LLM tinkerers, and anyone tracking the open vs closed capability gap.

## Try it

llm install llm-lmstudio && llm -m qwen3.6-35b-a3b 'Generate an SVG of a pelican riding a bicycle'

## Links

- [Blog Post](https://simonwillison.net/2026/Apr/16/qwen-beats-opus/)
- [Substack](https://simonw.substack.com/p/qwen36-35b-a3b-on-my-laptop-drew)
- [Qwen Tag](https://simonwillison.net/tags/qwen/)