maaza-nlm-orchestrator-9.6m-v1.2

maaza-nlm-orchestrator-9.6m-v1.2

63% adversarial accuracy (+37% from v1.0) · 86% in-distribution · 39ms latency · 9.60M parameters

The fastest adversarial-robust orchestrator ever shipped under 20M parameters. The official routing brain for the MCPBodega ecosystem.

---

What's New in v1.2

• +37% adversarial robustness (26% → 63%) via 10x upsampled adversarial training
• Trained on 496 diverse adversarial examples (typos, slang, synonyms, contractions)
• No wrapper, no fallback — pure model improvement
• Same architecture, same latency class

---

Performance (v1.2 — December 2025)

v1.2's adversarial training reduces the need for production wrappers. The model now handles typos, slang, and informal input natively.

---

Paper

Task-Specialized Micro Language Models Outperform Larger Zero-Shot Models on Structured Data Extraction

Authors: CycleCore Technologies Date: November 22, 2025 Version: 0.7

Full Paper (PDF)

---

NLM Taxonomy (CycleCore, 2025)

maaza-nlm-orchestrator-9.6m is the current flagship of the NLM category.

---

Model Card

Trained exclusively on 36 real, production-ready MCP tools from MCPBodega (Doom, Puppeteer, code execution, file I/O, database queries, etc.). No synthetic or placeholder tools.

---

Comparison

Ranks #1 under 20M parameters on adversarial robustness.

---

One-line deployment

mcpbodega deploy nano-orchestrator

---

Usage Example (PyTorch)

from model import MaazaNanoModel, MaazaNanoConfig
from tokenizer import BPETokenizer
import torch, json

tokenizer = BPETokenizer.load("tokenizer.json")
config = MaazaNanoConfig(**json.load(open("config.json")))
model = MaazaNanoModel(config)
model.load_state_dict(torch.load("model.pt", weights_only=True))
model.eval().cuda()

prompt = "<|user|>search for cats on the internet<|assistant|>"
input_ids = torch.tensor([tokenizer.encode(prompt)]).cuda()

with torch.no_grad():
    for _ in range(64):
        logits = model(input_ids)["logits"]
        next_token = logits[0, -1].argmax(-1)
        input_ids = torch.cat([input_ids, next_token[None, None]], dim=-1)
        if next_token.item() in tokenizer.special_tokens.values():
            break

print(tokenizer.decode(input_ids[0].tolist()))
# → [{"tool": "web_search", "params": {"query": "cats"}}]

---

v1.2 Training Details

Fine-tuned from v1.0 on adversarial data:

• 2,520 clean examples
• 496 adversarial examples (10x upsampled = 4,960)
• Total: 7,480 examples (66% adversarial ratio)
• 5 epochs, LR 3e-5, batch 32

Adversarial perturbations:

• Typos (random character swaps)
• Synonyms (search→find, weather→climate)
• Slang suffixes (pls, bruh, yo, lol)
• Contractions (you→u, please→plz, for→4)

---

Supported Tools (36)

---

License

Apache 2.0

---

Citation

@misc{cyclecore2025maaza-nlm-v1.2,
  author       = {CycleCore Technologies},
  title        = {maaza-nlm-orchestrator-9.6m-v1.2: Adversarial-Robust Tool Routing},
  year         = {2025},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/CycleCoreTechnologies/maaza-nlm-orchestrator-9.6m-v1.2}
}

---

CycleCore Technologies · @CycleCoreTech

cyclecore.ai · mcpbodega.com · slmbench.com

December 2025