Zyphra previews ZAYA1-74B on AMD

This is more infrastructure proof point than finished model drop: Zyphra is showing that large-scale pretraining on AMD is real, but the preview status means the benchmark story is still provisional.

• –The model is explicitly pre-RL and not instruction- or chat-tuned, so head-to-head benchmark claims need caution
• –The scale is substantial: roughly 15T pretraining tokens, 256k context extension, and an MoE design aimed at long-context efficiency
• –The AMD-only training stack matters for developers watching alternative GPU ecosystems, especially MI300X and Pensando networking
• –Community reaction is already skeptical about pass@4 vs pass@1 comparisons, so outside validation will matter more than the launch post
• –Apache 2.0 weights lower the friction for adoption if Zyphra follows through with the final RL-tuned model