Sabi open-sources 27B BCI model with 15% hallucination rate, far below GPT-6
lukas_m_ziegler · x · 2026-10-20
BCI startup Sabi is building a cap that turns imagined words into prompts, and released the model that sits on the other end: Sabi-Noema (27B, Apache 2.0, built on Qwen3.8-27B with RL on open-source environments only).
The job is brutal: EEG decodes only a noisy guess of a half-formed thought, so a confident wrong answer means putting words in the user's head they never thought. Hence the metrics:
- Hallucination rate 15.1% vs 30.4% for its own base and 85.3% for GPT-6 Luna
- Terminal-Bench 2.1: 76.0, best in their comparison set
- τ³-bench: 259/278 (next best 220); MCPMark: 48/55 (next best 42)
- 29.3K median output tokens per task vs 65.3K for the base
Half the hallucination rate and half the token spend of the model it was built from.
More from Models
- NVIDIA's Dory1.0-4B Looped Model Lets You Pick Inference Depth at Runtime — lmoroney · 2026-10-20
- OpenAI dropped 722 AI-generated math papers in a day; can humans read them? — songyoupeng · 2026-10-20
- JevBench adds 34 decision models in four days; Nimble-v4-max price cut to $0.04/1M tokens — airesearch12 · 2026-10-20
- Mercury 3 benchmarks: beats GPT-6 Luna and Haiku 5.5 at 1,068 tok/s — StefanoErmon · 2026-10-20
- Inception Labs releases Mercury 3, a diffusion LLM claiming 1,000+ tok/s at GPT-6-level quality — StefanoErmon · 2026-10-20
- OpenAI upgrades open-source devs to $200/mo ChatGPT Pro, free for six months — NielsRogge · 2026-10-20