Gensyn live-trains open-8b: every step auditable on your laptop, overhead down to 1.3x
benfielding · x · 2026-11-10
Gensyn has opened up the in-progress training run of open-8b, an 8B-parameter model. Every training step's state is committed as it happens, and anyone can replay a step on their own hardware for bitwise-exact comparison.
- open-1b proved auditable training is possible: public fingerprints per step, verifiable on everyday hardware — but replayability cost 5.5x the GPU-hours of a standard run.
- open-8b is 5x larger, trained on 2.5x the tokens, with audit overhead down to about 1.3x, aiming to show auditable models can be genuinely useful.
- Once post-training completes, it will serve as the judge model for Delphi, Gensyn's information-market platform, settling markets with real customer funds.
The team frames transparency as a ladder: open weights give you the cake, an open recipe adds the recipe card, an auditable model invites you into the kitchen.
More from Models
- GPT-6 Astra edits Qwen3.5-0.8B weights directly: 0.016% of params, MATH-500 from 7.9% to 48.1% — sebkrier · 2026-11-10
- Fable 5.2 launch offers: up to 3 usage resets for long-time subscribers, one-week trial for Pro — ClaudeDevs · 2026-11-10
- Fable 5.2 effort levels: API defaults to high, xhigh reserved for long-running agentic work — ClaudeDevs · 2026-11-10
- Anthropic extends preserved thinking to all API accounts on Fable 5.2 to block distillation — ClaudeDevs · 2026-11-10
- Matthew Berman: Fable 5.2 has 'completely solved' motion graphics — one-shot Fall Guys clone included — MatthewBerman · 2026-11-10
- Claude Fable 5.2 tested: one prompt turns 5GB of federal data into a US water-safety map — alliekmiller · 2026-11-10