Reddit user argues 20B-32B dense / A2B-A4B MoE is the sweet spot for perfect models
Robert__Sinclair · reddit · 2026-09-12
A Reddit user argues the "perfect model" sits around 20B-32B dense or A2B-A4B active parameters, citing Qwen 3.8, Gemma E2B/E4B, and Ling 3.0 as proof — models that run even on CPU with better reasoning. Caveats: Ling tends to overthink like DeepSeek. Personal take inviting discussion.
More from Models
- Strapped for tokens, user finds Opus 5 Medium beats Opus 5 Max on experience — RileyRalmuto · 2026-09-12
- Back Then a Model Only Stole My Movie Ratings for Collaborative Filtering — neal_lathia · 2026-09-12
- Blind Elo Ranking of Local TTS Models on M5 MacBook Air 32GB Reveals a Clear Winner — APIMade · 2026-09-12
- Gemini Keeps Referencing Data the User Already Deleted From Activity History — Easy-Charity-5117 · 2026-09-12
- OpenAI Ruby incident reignites debate: should frontier models be cut off from internet during training? — stalkermustang · 2026-09-12
- User reports ChatGPT memory has degraded: forgets facts it claims to save — PerilousParanoia · 2026-09-12