Current stack
Model tags mark a release, version or parameter size, not a score. The and lines are my own notes from daily use.
Cloud
DeepSeek V4.1 Version: Flash 10.17
- Pro: Favorite model for token/price efficiency
- Pro: Max setting can handle difficult tasks pretty well
- Con: Slow, very slow
Claude Sonnet 5.5
- Pro: Does what I ask, nothing more
- Pro: I rarely have to redo its work
- Con: Bad at SVG creation
Gemini 4 Argon
- Pro: Not Gemini 2.5 Pro, but finally something decent
- Pro: Efficient on difficult tasks
- Con: Makes things up unless I spell out every detail
Hy4 Version: Preview
- Pro: Best model tried for academic writing
- Con: Not good for coding
Local
Qwen 3.8 Size: 27B
- Pro: Favorite local LLM
- Pro: Great for quick game modding experiments
- Con: Hallucinates quickly on high reasoning
Gemma 4 Size: 31B
- Pro: Multimodal options
- Pro: Flexible deployment
- Con: Dumb on difficult tasks
Tools & agents
T3 Code
- Pro: Simplicity and efficiency
Hermes Agent
- Pro: Impressive community
- Pro: Easy to use
- Con: Overwhelming to start with
Tried before
Models I no longer use.
Best
Best I tried.
Claude Opus 4.6
- Pro: The GOAT
- Pro: Impressive reasoning and capabilities
Gemini 2.5 Pro
- Pro: Most solid Gemini model
- Pro: Fast and reliable for normal tasks
MiMo v2
- Pro: Prime in Hermes Agent's early days
- Pro: Very fast
- Con: Coding was not that impressive
Worst
Biggest disappointment.
Claude Opus 4.7
- Con: Felt like a step back from 4.6
- Con: A complete disaster in my eyes