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

2026-04-16 · Source: Simon Willison's Weblog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Intermediate, quick

Summary

On April 16, 2026, a comparison of image generation capabilities between Alibaba's Qwen3.6-35B-A3B and Anthropic's Claude Opus 4.7 revealed unexpected results. Using a 20.9GB quantized version of Qwen3.6-35B-A3B-UD-Q4_K_S.gguf running locally on a MacBook Pro M5 via LM Studio, the Qwen model produced a superior image of a "pelican riding a bicycle" compared to Claude Opus 4.7, which struggled with the bicycle's frame. A subsequent test involving a "flamingo riding a unicycle" also favored Qwen, which generated a more charismatic and detailed SVG. Despite the "pelican benchmark" being a humorous, informal test, its historical correlation with general model utility appears to be breaking, as a smaller, locally run model outperformed a major proprietary release.

Key takeaway

For AI engineers evaluating image generation models, do not solely rely on general utility benchmarks. Your teams should consider testing smaller, quantized models like Qwen3.6-35B-A3B for specific creative outputs, especially when local inference is a priority. This approach might yield superior results for niche tasks compared to larger, proprietary models, challenging assumptions about model hierarchy.

Key insights

Local, quantized models can surprisingly outperform larger proprietary models in specific creative tasks.

Principles

• Model utility does not always correlate with benchmark performance.
• Quantized models offer significant local inference capabilities.

In practice

• Test local LLMs for niche creative tasks.
• Explore quantized models for specific image generation needs.

Topics

• Qwen3.6-35B-A3B
• Claude Opus 4.7
• Image Generation
• Model Benchmarking
• Quantized Models

Code references

• agustif/llm-lmstudio

Best for: AI Engineer, Computer Vision Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, Prompt Engineer

Related on AIssential

• Granite 4.1 3B SVG Pelican Gallery — Quantization level does not guarantee improved output quality for specific creative tasks.
• not much happened today — AI models are advancing in design, coding, and agentic capabilities, with a focus on efficiency, local deployment, and robust evaluation.
• Qwen3.6-27B: Flagship-Level Coding in a 27B Dense Model — Qwen3.6-27B offers flagship coding performance in a significantly smaller, dense model.
• Open image generation models are closer to closed-source quality than this sub thinks [D] — Open-source image generation models now rival closed-source APIs in key performance areas like compositional control and text rendering.
• DirectorBench: Diagnosing Long-Form Video Generation with Personalized Multi-Agent Evaluation — Long-form video generation requires diagnostic, profile-aware evaluation beyond aggregate scores to identify specific bottlenecks.
• New benchmark confirms AI video generators look stunning but still can't reason about the world — Current AI video generators excel visually but lack fundamental world understanding and logical reasoning capabilities.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Simon Willison's Weblog.