ZAYA1 Models by Zyphra

ZyphraApache 2.0Open source
Zyphra releases · 1 in the last 12 monthsChangelog →
1 model2026Up to 32K ctx

Last refreshed 2026-05-12. Next refresh: weekly.

Details

ResearcherZyphra
LicenseApache 2.0OSI-approved
Commercial usePermitted
Models1
Released2026
Max context32K

Capabilities

ReasoningAll models

Links

WebsiteHuggingFace

About

ZAYA1 is Zyphra's family of mixture-of-experts (MoE) language models. Zyphra says ZAYA1 was trained entirely on AMD's hardware, software, and networking stack. ZAYA1-8B, the post-trained reasoning model, has 8.4B total parameters with about 760M active per token and is aimed at detailed long-form reasoning, especially math and coding; its small size also suits on-device, local use. The line also includes the ZAYA1-base and ZAYA1-reasoning-base pretraining checkpoints, ZAYA1-VL-8B, a vision-language model built on the 8B model, and ZAYA1-74B-Preview, a 74B-total, 4B-active reasoning base that has not been tuned for chat. Weights are open on Hugging Face.

Decision facts

Best fit
reasoningcodingmath-heavy prompts
Capability starting point
ZAYA1-8B with 32K context and reasoning
Lowest tracked input
Not tracked
Closest related family
Zamba

Current Variants

Use-when guidance is based on each model's tracked capabilities, context window, release date, and replacement status.

1 in view
ZAYA1-8BCurrent

Use when the workload needs reasoning, 32K context, and 8.4B parameters.

2026-05reasoning32K context8.4B parameters

Release Timeline

1 release group
2026-05
1 current
ZAYA1-8B
reasoning32K context8.4B parameters
Current

Specifications(1 models)

ZAYA1 model specifications comparison
ModelReleasedContextParametersReasoning
ZAYA1-8B2026-0532K8.4BYes