RedNote AI lab's TEMPO RL method: 16B MoE scores >30% on ARC-AGI 3
GregKamradt · x · 2026-08-15
According to a retweet, RedNote's AI lab announced a new RL regimen (TEMPO, beta version) that rewards world exploration, allowing a 16B active MoE model to score >30% on ARC-AGI 3. The model, dots3-note Preview, is an interim preview release with reinforcement learning not yet complete.
More from Models
- Anthropic uses internal model far better than Mythos 5, no release planned — kimmonismus · 2026-08-15
- Test shows Qwen3.8-27B achieves high precision in long video timestamping — vanstriendaniel · 2026-08-15
- RTX 3090 gets 35 t/s on Qwen 3.8 27B — cviperr33 · 2026-08-15
- Gemini 3.7 Flash Launches with Enhanced Reasoning and Tool Use — arvind_io · 2026-08-15
- Anthropic confirms Mythos 5 as top internal model, hints at mysterious Model 2 — scaling01 · 2026-08-15
- Users report random stop behavior in Qwen 2.5/3 during long context generation — T_rex2700 · 2026-08-15