Willison's head-to-head comparison demonstrates that Qwen3.6-35B-A3B, activating only ~3B parameters on consumer hardware, produced a visually superior SVG pelican to Claude Opus 4.7 via API. His systematic documentation of local model progress and willingness to declare clear winners gives this comparison particular weight in the developer community.
Willison chose SVG generation deliberately — it produces structured code output with a visual result that's trivially easy to evaluate. Unlike benchmark gaming or vibes-based scoring, you simply look at two pelicans and one is clearly better, making the comparison unusually credible.
The editorial highlights that this test avoids the typical pitfalls of model comparisons — no benchmark gaming, no subjective scoring. The task's visual clarity is what made it resonate on Hacker News, earning 372 points.
The editorial argues that MoE is no longer experimental — it has been validated by multiple major labs shipping production-quality models in 2025-2026. The architecture breaks the old equation where bigger models required bigger GPUs, enabling a 35B-parameter model to run with the footprint of a 3B dense model.
Simon Willison — developer, blogger, and one of the more credible voices in the applied-AI space — published a head-to-head comparison on April 16th that caught Hacker News's attention (372 points and climbing). The test: ask both Qwen3.6-35B-A3B running locally on his laptop and Anthropic's Claude Opus 4.7 via API to draw a pelican. The local model won.
The specifics matter. Qwen3.6-35B-A3B is Alibaba's latest mixture-of-experts (MoE) release — 35 billion total parameters, but only approximately 3 billion active on any given token. That architecture is what makes it runnable on a laptop at all. A model activating just 3B parameters on consumer hardware produced visually superior output to one of the most capable cloud models in production. The task was SVG generation — structured code output with a visual result that's trivially easy to evaluate. No benchmark gaming, no vibes-based scoring. You look at two pelicans and one is better.
This isn't Willison's first rodeo with local models. He's been systematically documenting their progress through his LLM tool and blog, and his willingness to name winners and losers — rather than hedging with "it depends" — is precisely why this particular comparison resonated.
Three trends are converging, and this pelican is their mascot.
First: MoE architectures have changed the local-model calculus. The old mental model was simple — bigger model means better output, and bigger models need bigger GPUs. MoE breaks that equation. By routing each token through a subset of expert networks, a 35B-parameter model can run with the memory footprint and compute cost of a 3B dense model. The quality-per-watt ratio of MoE models has improved roughly 10x in the past 18 months, and Qwen3.6 is the latest proof point. Alibaba, Mistral, and DeepSeek have all shipped production-quality MoE models in 2025-2026; this architecture is no longer experimental.
Second: the task matters as much as the model. SVG generation is a constrained creative task — the output is code, the vocabulary is limited, and the quality is immediately verifiable. For these kinds of structured-output tasks, local models don't need to match a cloud model's general reasoning ability. They just need to be good enough at the specific thing you're asking. On constrained, code-adjacent tasks like SVG/HTML/CSS generation, the gap between local and cloud models has effectively closed. This doesn't mean Qwen3.6-35B-A3B will match Opus on multi-step reasoning chains or 100K-context analysis — it almost certainly won't. But for a surprisingly large category of daily developer tasks, "good enough" arrived on your laptop.
Third: the economics are asymmetric and getting more so. Every Opus API call costs money. Every local inference costs electricity. For a developer running dozens of creative generation tasks per day — generating icons, writing boilerplate, producing test data — the difference compounds. At scale, free-per-marginal-query changes behavior. Developers who wouldn't bother with an API call for a throwaway task will absolutely hit a local model. That expanded usage surface is where local models create value that cloud models structurally can't.
The Hacker News discussion predictably split between two camps. The "local-model maximalists" see this as vindication — further proof that running your own models is the future. The pragmatists correctly note that a single SVG generation task doesn't invalidate Claude's advantages on complex reasoning, long-context work, or tool use. Both camps are right, and that's exactly the point. The question is no longer "are local models good enough?" but "good enough for which tasks?" — and the answer keeps expanding.
It's also worth noting the source: Alibaba's Qwen team has been shipping at a relentless pace. Qwen3.6 follows the Qwen3 release from earlier in 2026, with the A3B variant specifically targeting the local-inference market. This is deliberate product strategy — Alibaba can't compete with Anthropic or OpenAI on cloud API distribution, but they can win the local-inference developer mindshare by releasing models optimized for consumer hardware. The competitive dynamics here are genuinely interesting: the Chinese labs are building their moat in open weights while US labs build theirs in proprietary APIs.
If you're a developer who uses AI for code-adjacent creative tasks — generating SVGs, HTML mockups, CSS animations, test fixtures, config files — you should be running a local MoE model right now. The tooling has matured: Ollama, llama.cpp, and LM Studio all support Qwen3.6-35B-A3B, and setup is measured in minutes, not hours.
The practical architecture that's emerging is a tiered model stack: local MoE models for high-frequency, low-complexity tasks (generation, formatting, boilerplate); cloud APIs for high-complexity, low-frequency tasks (architecture review, multi-file refactoring, long-context analysis). Developers who adopt this tiered approach will spend significantly less on API costs while maintaining quality where it counts. The key is knowing which tier a task belongs to — and Willison's pelican test is a useful heuristic: if the output is constrained, verifiable, and code-shaped, try local first.
One caveat: don't over-index on a single comparison. SVG generation is a sweet spot for local models — structured output, limited token count, visual verification. Your mileage will vary on open-ended prose, complex reasoning, or tasks requiring broad world knowledge. The smart move is to run your own comparisons on your actual workloads, not to declare victory based on pelican art.
The trajectory is clear even if the timeline isn't. MoE architectures will continue to improve the quality-per-parameter ratio. Hardware will continue to get faster and cheaper. Within 12-18 months, expect laptop-runnable models that match today's cloud frontier on most practical coding tasks. The cloud API providers know this — which is why Anthropic, OpenAI, and Google are all investing heavily in agentic workflows, tool use, and long-context capabilities that are structurally harder to replicate locally. The future isn't local vs. cloud; it's a portfolio allocation problem, and the local allocation just got a bigger slice.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.