Håvard Ihle
@htihle.bsky.social
AI researcher, former cosmologist. https://htihle.github.io/
Qwen 3 coder next (80b3a) scores 34.4% on WeirdML, which is pretty good for it's size, especially for a non-reasoning model.
Probably a good choice for agentic coding if you need a small local model.
For more info see: htihle.github.io/weirdml.html
Probably a good choice for agentic coding if you need a small local model.
For more info see: htihle.github.io/weirdml.html
February 4, 2026 at 1:02 PM
Qwen 3 coder next (80b3a) scores 34.4% on WeirdML, which is pretty good for it's size, especially for a non-reasoning model.
Probably a good choice for agentic coding if you need a small local model.
For more info see: htihle.github.io/weirdml.html
Probably a good choice for agentic coding if you need a small local model.
For more info see: htihle.github.io/weirdml.html
kimi-k2.5 scores 45.6% on WeirdML, up from 42.8% for k2-thinking. This is better than claude opus 4, and close behind sonnet 4.5 and deepseek-3.2-speciale at 47.7 and 46.7%, but way behind gpt-5.2 at 72.2% and the other leading closed models.
For more details: htihle.github.io/weirdml.html
For more details: htihle.github.io/weirdml.html
January 30, 2026 at 6:24 PM
kimi-k2.5 scores 45.6% on WeirdML, up from 42.8% for k2-thinking. This is better than claude opus 4, and close behind sonnet 4.5 and deepseek-3.2-speciale at 47.7 and 46.7%, but way behind gpt-5.2 at 72.2% and the other leading closed models.
For more details: htihle.github.io/weirdml.html
For more details: htihle.github.io/weirdml.html