9B Model Trained on Bittensor Scores 43.6% on SWE-bench, Beating Qwen3.5-9B
const_reborn · x · 2026-11-05
Reliquary AI released Teutonic-I-Reliquary, a 9B model pre-trained on Teutonic (3) and post-trained entirely on Bittensor using untrusted GPUs ranging from RTX 3090s to B200s. It scores 43.6% on SWE-bench Verified versus Qwen3.5-9B's 35.1%. Observers say the notable part isn't the size but the alignment/competitive-training approach that punched through the SOTA weight class.
Related event: 9B Coding Model Trained on Bittensor Beats Qwen Flagship on SWE-bench(2 posts)→
More from coding & agent
- 5 Grok bots chase overdue freelance invoices: a 5-agent workflow from a 30-day build series — nima_owji · 2026-11-06
- Zerg AI launches ZTC, an agent client built for mission-critical work — idanbeck · 2026-11-06
- ZTC architecture: TypeScript agent plus Rust screen layer over local JSON protocol — idanbeck · 2026-11-06
- Firecrawl launches free life-science research index with 41M+ papers for AI agents, raises $75M Series B — devdigest · 2026-11-06
- SWE-rebench V2.1 RL ships 13,337 free RL environments for coding agents from real GitHub issues — OfirPress · 2026-11-06
- Monologue's Playbook: When to Adapt Your Own AI Model — every · 2026-11-06