ZAYA1-8B
by Zyphra
ZAYA1-8B is a reasoning-focused mixture-of-experts (MoE) language model from Zyphra with 8.4 billion total parameters and only 760 million active parameters per token. It was trained entirely on AMD Instinct MI300X hardware and achieves competitive reasoning performance at a fraction of the compute cost of larger models.
Large Language Models
Proprietary
Primary Use Cases
- Mathematical reasoning
- Code generation and analysis
- Puzzle solving
- Cost-efficient inference at scale
- On-premise deployment without NVIDIA hardware
✅ Strengths
- Competitive reasoning performance vs. much larger models
- Extremely efficient: only 760M active parameters per token
- Open-source under Apache-2.0 license
- Demonstrates AMD hardware viability for frontier model training
- Novel Markovian RSA enables extended reasoning with constant memory footprint
⚠️ Limitations
- 8.4B total parameters limits absolute capability ceiling vs. frontier models
- MoE architecture may have higher memory requirements than dense models of similar active size
- Relatively new model with limited third-party benchmark validation
- Primarily reasoning-focused; general instruction following may lag larger models
🏆 Performance Benchmarks
Standardized benchmark scores for objective comparison
Reasoning vs. larger models(Reasoning)
0qualitativeCompetes with much larger models on reasoning tasks while activating only ~760M parameters per token. Trained on AMD hardware.
Technical Specifications
Developer
Zyphra
Category
Large Language Models
Type
🧠 Foundation Model
License
UNKNOWN
Performance Benchmarks
Reasoning vs. larger models0qualitative
Source: View benchmark details
Trust & Privacy
Health Status
🟢 Active
Privacy Grade
🛡️ Grade A
Trains on User Data
✅ Does not train
Certifications
None listed
Ecosystem & Integrations
API Access
❌ No
Chrome Extension
❌ No
Mobile App
❌ No
Pricing Model
🔓 Open Source
Free Tier
✅ Available
Pricing Breakdown
Open Source
Free to download and self-host under Apache-2.0 license
0