NewsHive
CONTACT USANALYST PORTAL →
aiBACKGROUND

AI Models Drawing Pelicans

BACKGROUND
4 SIGNALSFIRST DETECTED 24 July 2026UPDATED 15 October 2026
The NewsHive View

On 24 Jul, a LocalLLaMA post compared SVG generation for a pelican riding a bicycle across Laguna S2.1 118B Q2 LX, Gemma4 12b Q8_0, and Qwen 3.6 35B A3B IQ4 +Q8_0. On 14 Aug, a user said Qwen3.8-27B responded to a request to draw a pelican by producing an entire SVG animation. Another Qwen3.8-27B pelican post followed that day. By 13 Sep, attention had moved to “Qwen3.8 flash next” and a claim of untrained SVG generation. The informal pelican test has become a recurring demonstration of how these models produce vector graphics. Apparently, the bird is now infrastructure.

Even if factual, the impact is 0%. These LocalLLaMA demonstrations give users examples to inspect. They do not change model availability, pricing, training processes, or documented capabilities.

Watch for reproducible prompts, generated SVG files, and controlled comparisons across the named models. Official documentation addressing the “untrained SVG generation” claim would materially change the picture.

How the story developed
13 Sept
14 Aug
yet another Qwen3.8-27B pelican post
OPINION · WHAT · LocalLLaMA
Other feeds and industry voices
24 Mar
Distilled qwen 3.5 27b is surprisingly good at driving Cursor.

<!-- SC_OFF --><div class="md"><p>I'm using this <a href="https://huggingface.co/Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2-GGUF">opus 4.6 distilled version of

OPINION · HOW · LocalLLaMAOPEN STORY
30 Apr
Qwen-Scope: Official Sparse Autoencoders (SAEs) for Qwen 3.5 models

<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1szrbub/qwenscope_official_sparse_autoencoders_saes_for/"> <img alt="Qwen-Scope: Official Sparse

OPINION · WHAT · LocalLLaMAOPEN STORY
26 Apr
Showed 4 AI models some abstract Kandinsky-style Pokémon art with no hints, the results are kind of insane

<!-- SC_OFF --><div class="md"><p>found this artist &quot;8th Project&quot; on instagram who drew Pokémon as pure geometric abstraction, decided to just drop it on every Al I had

OPINION · WHO · ArtificialInteligenceOPEN STORY
1 May
Tested the new Claude MCP that runs 30+ image and video models in one chat. 50 minutes vs 2.5 hours on the same brief

<!-- SC_OFF --><div class="md"><p>Until last week, generating an image inside Claude meant Claude wrote you a prompt. Then you copied it. Opened another tab. Pasted it into

OPINION · WHAT · ArtificialInteligenceOPEN STORY
10 May
Anybody else noticing how good gemma-4-26b-a4b is with one-shotting three.js?

<!-- SC_OFF --><div class="md"><p>I wrote up this little python app to cycle through a bunch of prompts like this:</p> <table><thead> <tr> <th align="left">Single HTML file using

OPINION · HOW · LocalLLaMAOPEN STORY
22 Apr
The new Qwen3.6-27B just gave me definitely the best pelican riding a bicycle I've had from a 16.8GB model file! simonwillison.net/2026/Apr/22/...

The new Qwen3.6-27B just gave me definitely the best pelican riding a bicycle I've had from a 16.8GB model file! simonwillison.net/2026/Apr/22/...

OPINION · WHAT · Simon WillisonOPEN STORY
14 May
Anthropic merges consecutive same-role messages, OpenAI doesn't (+4 tokens), anyone token-counted this on open-weight models?

<!-- SC_OFF --><div class="md"><p>I build context/harness optimization tooling, so provider-side serialization quirks actually matter to me. If you're optimizing over prompts, you

OPINION · WHAT · LocalLLaMAOPEN STORY
13 May
qwen3.6 just stops

<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1tbzr64/qwen36_just_stops/"> <img alt="qwen3.6 just stops"

OPINION · WHAT · LocalLLaMAOPEN STORY
Sources
LocalLLaMA×4

NewsHive monitors these sources continuously. All signal titles above link to the original reporting.

Intelligence by NewsHive. Need help navigating what this means for your business? Contact GeekyBee →