Post-trained Qwen3-4B tops Jane Street's MegaGem, beating GPT-5.5
tokenbender · x · 2026-08-14
A developer shares how post-training Qwen3-4B-Instruct made it rank #1 on Jane Street's auction game MegaGem, surpassing GPT-5.5 and Claude Opus 4.8, but self-play RL wasn't the key.
More from Models
- Pestle 27B Ternary: first medical ternary model, Apache 2.0 — Individual-Dot5488 · 2026-08-15
- Frontier model prices drop across the board, open and closed source, making agents cheaper — markjeffrey · 2026-08-15
- RedNote AI lab's TEMPO RL method: 16B MoE scores >30% on ARC-AGI 3 — GregKamradt · 2026-08-15
- Orion-16B passes 100B tokens, largest LLM pretrained with decentralized compute — const_reborn · 2026-08-15
- AI model benchmarks: baseline crushed, top models nearly finish course — const_reborn · 2026-08-15
- Qwen3.8 gets Day-0 support from LightSeek, boosting inference performance by 30%+ — Alibaba_Qwen · 2026-08-15