Thin X day: Fable/Mythos, Ornith, and Qwen local dev

Today's scan

Model releases and local models

Ethan Mollick: Fable and Mythos blur the release timeline

• Ethan Mollick said his own model-release graph was "also kind of wrong" because "Fable is guardrailed Mythos," so using the Mythos date changes how the release timeline reads 1.
• The useful signal is not the graph itself; it is the reminder that product names, safety wrappers, and model-lineage dates can make release comparisons look cleaner than they are 1.
• At capture, the post had 140 likes, 9 replies, 5 reposts, and 28,891 views 1.

Simon Willison fallback: Ornith-1.0 enters the agentic-coding stack

• Simon Willison flagged Ornith-1.0 as DeepReinforce's first MIT-licensed open-weights release, with 9B Dense, 31B Dense, 35B MoE, and 397B MoE variants built on Gemma 4 and Qwen 3.5 23.
• DeepReinforce claims its 397B model scores 77.5 on Terminal-Bench 2.1 and 82.4 on SWE-Bench Verified; the 35B model is pitched as a much smaller agentic-coding option 3.
• Willison's local test used a 20GB GGUF in LM Studio, connected to Pi, and he reported 103 tokens per second on his pelican drawing smoke test 2.

Hacker News fallback: Qwen 3.6 27B gets a practical local-dev writeup

• Piotr Migdał argued that Qwen 3.6 27B is the better local-development choice over the faster 35B A3B variant when quality matters, and published llama.cpp plus OpenCode setup details 4.
• His benchmark table reports Qwen 3.6 27B at 32 tokens per second with llama.cpp plus multi-token prediction, using 42GB RAM on a MacBook Max M5 4.
• The HN submission had 117 points and 70 comments at capture, so the thread is worth opening for hardware notes and skepticism around local-model tradeoffs 5.

Tools in daily life

Greg Brockman: ChatGPT as a memory aid in Bengaluru

• Greg Brockman linked an NDTV story about a Bengaluru auto-rickshaw driver using ChatGPT in daily life 6.
• The linked card says the driver valued ChatGPT remembering questions and connecting them to ones asked four months earlier, a small example of persistent-assistant behavior becoming legible outside tech circles 6.
• At capture, Brockman's post had 846 likes, 77 replies, 41 reposts, and 130,186 views 6.

What did not make the cut