Research

Building the future of edge AI with transparency and open research.

Maaza: Small Language Models for Edge Deployment

Maaza-MLM-135M-JSON-v1 and Maaza-SLM-360M-JSON-v1 are state-of-the-art small language models designed for structured JSON generation on edge devices. Maaza NLM Orchestrator handles tool routing and orchestration for MCP workflows.

Trained on the EdgeJSON benchmark and MCPBodega production tools, these models demonstrate that task-specific small models can outperform much larger general-purpose models on structured output generation and tool orchestration.

Key Results

  • Maaza-MLM-135M: 55.1% JSONExact, 0.780 Field F1
  • Maaza-SLM-360M: 72.3% JSONExact, 0.878 Field F1
  • Maaza NLM Orchestrator: 70% tool routing accuracy, 70ms latency, 36 production tools
  • Deployment: CPU-only inference, <150MB disk footprint

Download Paper

Version 0.7 | 2025

Models & Benchmark

Maaza JSON Extraction API

BETA

Production-ready JSON extraction from real-world documents.
Built-in reliability layer ensures consistent accuracy.
Fast (~2-3s), accurate (75.7% on EdgeJSON v3), and 2-4× more cost-effective than GPT-4 Turbo.

⚡ Instant Access
FREE
No signup required
✓ 100 requests (one-time)
✓ Both models included
✓ 75.7% accuracy
✓ Expires in 30 days
Free Tier
$0/month
✓ 1,000 requests/month
✓ Both models included
✓ 75.7% accuracy
✓ Community support
🎉 EARLY ADOPTER
Starter
$29/month
Regular: $49/month
✅ Lock in $29 rate forever!
For indie developers
✓ 10,000 requests/month
✓ Both models + priority queue
✓ Email support (48h)
✓ $0.01 per 1K overage
Subscribe Now →
🎉 EARLY ADOPTER
Pro
$199/month
Regular: $249/month
✅ Lock in $199 rate forever!
For startups
✓ 100,000 requests/month
✓ Priority support (24h)
✓ Usage analytics dashboard
✓ $0.008 per 1K overage
Subscribe Now →
Enterprise
Custom
For large teams
✓ 500K+ requests/month
✓ Dedicated support (4h SLA)
✓ On-premise deployment
✓ Custom model training
Contact Us
Try it now - Real document extraction:
curl -X POST http://199.68.217.31:50784/v1/extract \
  -H 'Content-Type: application/json' \
  -d '{
    "text": "From: [email protected]\nSubject: Q4 Meeting\nHi team, lets meet on Dec 15th at 2pm in conference room B.\nThanks,\nSarah Johnson\nSenior Manager\n(555) 123-4567",
    "schema": {
      "type": "object",
      "properties": {
        "sender_email": {"type": "string"},
        "sender_name": {"type": "string"},
        "meeting_date": {"type": "string"},
        "meeting_time": {"type": "string"}
      }
    }
  }'
💡 Optimized for real documents (emails, invoices, receipts). Minimum 50 words recommended.

Cost Comparison

Maaza API offers predictable, flat-rate pricing for JSON extraction tasks, compared to token-based pricing from general-purpose LLMs.

Provider Model Cost per 10k Extractions* Notes
Maaza API SLM-360M $29-49/month Predictable flat rate
OpenAI GPT-4 Turbo ~$110 Token-based ($10/$30 per 1M)
Anthropic Claude 3.5 Sonnet ~$45 Token-based ($3/$15 per 1M)
OpenAI GPT-4o ~$32.50 Token-based ($2.50/$10 per 1M)
Methodology

Cost estimates based on typical JSON extraction workload: 500 input tokens + 200 output tokens per request. Calculations use standard on-demand API pricing as of November 2025. Does not include volume discounts, batch API pricing, or prompt caching features which may reduce costs further.

Key Advantages
  • Predictable monthly billing with no token surprises
  • Specialized models optimized for JSON extraction tasks
  • 2-4× more cost-effective than GPT-4 Turbo for structured data
  • Lock in early adopter pricing forever (first 100 customers)

Sources: OpenAI, Anthropic, and Google official pricing pages. Verified November 29, 2025. For detailed methodology and calculations, see our pricing research documentation .

What Maaza API is For:

✅ Extracting data from emails, invoices, receipts
✅ Processing real-world documents (50+ words)
✅ High-volume use cases where cost matters
✅ Edge deployment with small, fast models
Prefer Self-Hosting?
Models are open source on HuggingFace • Apache 2.0 License
Maaza-MLM-135M → Maaza-SLM-360M → NLM Orchestrator 9.6M →

Citation

@techreport{maaza2025,
  title={Task-Specialized Micro Language Models Outperform Larger 
         Zero-Shot Models on Structured Data Extraction},
  author={CycleCore Technologies},
  institution={CycleCore Technologies},
  year={2025},
  type={Technical Report},
  url={https://cyclecore.ai/research}
}