Zyphra's ZAYA1-8B Matches Frontier Models on Math With Under 1B Active Parameters
What Happened
Zyphra released ZAYA1-8B on May 7, 2026, a mixture-of-experts reasoning model that matches or exceeds substantially larger open-weight models on math, coding, and reasoning tasks. Using novel "Markovian RSA" test-time compute, ZAYA1-8B approaches frontier models including Claude 4.5 Sonnet, Gemini 2.5 Pro, and DeepSeek v3.2 on math benchmarks while running on a fraction of the inference cost.
My Take
The "frontier vs open" gap is closing on a per-task basis faster than the labs want to admit. Math and code are the easiest domains because they have automatic verifiers — expect customer service, legal research, and document review to follow within 12 months. The strategic implication is brutal for the big labs: their pricing power on bounded, verifiable tasks is evaporating, and that's exactly the slice enterprises want to automate first. Frontier labs will retreat to agentic, multi-step, ambiguous work where evaluation is harder. That's where the real money is anyway.
Read Original Source