BeijingPrism scores 26 answer letters, not 248,000 tokens.
Almost simultaneously with DeepSeek’s DeepInfer release, 5-month-old StartLux open-sourced Prism on 22 October.
Decision models return probabilities over fixed answers, so Prism skips the full vocabulary and shares one document’s computation across several questions.
In StartLux’s own tests the speedup falls to 1.2 times on questions of about 8,192 tokens.
StartLux published its test configurations on GitHub, so outside teams can now check the 5.3 times claim.
How each outlet framed it
- Zhidx leans favorable
- frames decision models as a distinct inference niche and quantifies Prism's 2.8x and 5.3x speedups on a single H200
Sources: Zhidx