Bittensor SN81 generated ~3B tokens in a week, boosting Qwen3-4B math score from 37 to 73
bittingthembits · x · 2026-10-03
- It works: Subnet 81 (Reliquary) is an open market for post-training data. Permissionless miners trained Reliquary-4B with 6.6M rollouts and 12,852 RL updates. vs Qwen3-4B-Base under the same sampling: MATH-500 37.2→72.6, AMC23 15.6→47.5, HumanEval+ 52.1→72.9, MBPP+ 43.7→67.6. Weights and technical report are public.
- Still climbing: on a frozen test suite, the live run scores 74.5% on math first try (+34.3 over base) and 76.9% on code (+20.7).
- A data factory: a new batch-generation task added a week ago produced nearly 3B tokens with Qwen3.8-27B, spot-checked against model weights, now being distilled into Bittensor's Teutonic model; public access coming Tuesday.
More from Infra
- Sentdex straps 4x RTX Pro 6000 cards onto an Asus GB300 DGX Station — Sentdex · 2026-10-03
- Qwen 3.8 Flash Next q2_0 hits 10 tok/s on a 6GB VRAM laptop via Strata engine — dampflokfreund · 2026-10-03
- Strata hits 5-10x prefill speedup in two weeks, shaming llama.cpp maintainers — Training_Visual6159 · 2026-10-03
- One RTX 4080, One Night: Full AI Short Film With Qwen Images and Minimax H3 Video — CupQuakeBE · 2026-10-03
- Data centers are unpopular across America, but Loudoun County, their capital, thrives — arian_ghashghai · 2026-10-03
- Huawei's LogicFolding Tau design and Mate 90: China's chip progress still awaits independent testing — ingliguori · 2026-10-03