JamesQuartz/aenea-pinta-1.2-mini

AENEA Pinta-1.2 Mini Stable

50M-parameter semantic router for edge dispatch. Pinta-1.2 Mini Stable is the production-ready successor to the Pinta-1.1 Mini Beta, delivering 91.51% strict accuracy - a massive +19 point improvement over its predecessor - while maintaining the same 50M parameter budget and sub-15 ms CPU inference on commodity hardware.

Designed for always-on classification in agent swarms, CLI sidecars, and edge dispatch where latency matters more than peak accuracy.

Stable release notice. This is a Stable release. The routing head was trained on a strictly balanced 58,220-record dataset (1.46x max/min ratio across all 9 domains). Unlike the Beta, no calibration bias vector is required at inference time. The model can be used directly with raw softmax over the 9 domain logits.

---

What is this?

Pinta-1.2 Mini is not a generative language model. It is a semantic router that classifies incoming prompts into one of nine dispatch domains and emits a routing token that downstream systems can use to select the appropriate expert model or tool:

Class	Token	Dispatch to
SyntaxDevOps	<|reserved_23|>	Template engines, formatters, regex, data conversion
Code	<|reserved_24|>	Code-specialized models, debuggers, linters
Creative	<|reserved_25|>	Creative writing models, storytelling engines
RAG	<|reserved_26|>	Retrieval-augmented pipelines, document QA
Architecture	<|reserved_27|>	System design experts, DevOps, cloud infra
Math	<|reserved_28|>	Math/reasoning models, calculators, CAS
Knowledge	<|reserved_29|>	General knowledge / fallback models
Compliance	<|reserved_30|>	Legal domain models, policy engines, compliance review
Chat	<|reserved_31|>	Chat-optimized models, conversational agents

The router is trained to predict which token a full generative model would emit next, but does so in a single forward pass without generating any text.

---

Architecture

Component	Specification
Parameters	50M
Layers	20 (Cartan/Cittern)
Hidden size	512
Attention heads	8
Orthogonal attention	Block Householder, 8 reflections, block_size 64
Vocabulary	9,000 (QT-Cittern-1.0 tokenizer)
Context window	2,048 tokens
Precision	bf16 / fp16 / ONNX fp32 & fp16
Routing head	Causal [B, S, V] with 9 reserved token slice

---

Benchmarks

Evaluated on the Balanced Routing Benchmark v6 (1,060 prompts across 9 domains, zero train overlap, English-only):

Metric	Pinta-1.1 (226M)	Pinta-1.1 Mini Beta	Pinta-1.2 Mini Stable
Strict accuracy	79.22%	72.50%	91.51%
Macro F1	-	-	0.91
Weighted F1	-	-	0.92
Latency (p50, CPU)	~45 ms	~14 ms	~14 ms
Latency (p95, CPU)	~52 ms	~16 ms	~18 ms
Model size (ONNX fp32)	452 MB	210 MB	199.9 MB
Model size (ONNX fp16)	-	-	105.3 MB

Pinta-1.2 Mini outperforms the full 226M Pinta-1.1 by 12+ percentage points while remaining smaller and 3x faster.

Per-Domain Performance

Domain	Precision	Recall	F1
SYNTAX_DEV_OPS	0.96	0.73	0.83
CODE	0.80	0.86	0.83
CREATIVE_PROSE	0.96	0.83	0.89
LONG_CONTEXT_RAG	0.98	0.93	0.96
SYSTEM_CODE_ARCH	0.90	0.95	0.93
DENSE_STEM_MATH	0.94	0.99	0.96
GENERAL_KNOWLEDGE	0.76	0.97	0.85
HIGH_RISK_COMPLIANCE	1.00	0.97	0.99
CONVERSATIONAL_CHAT	0.98	0.97	0.98

---

Quickstart

Python Inference (ONNX)

import numpy as np
import onnxruntime as ort
from transformers import AutoTokenizer

# Load artifacts
tokenizer = AutoTokenizer.from_pretrained("tokenizers/QT-Cittern-1.0")
session = ort.InferenceSession("pinta-1.2-mini.fp16.opset18.onnx")

DOMAINS = [f"<|reserved_{i}|>" for i in range(23, 32)]
domain_ids = [tokenizer.convert_tokens_to_ids(d) for d in DOMAINS]

def route(prompt: str) -> dict:
    formatted = f"User: {prompt}\nAssistant: "
    ids = tokenizer.encode(formatted, return_tensors="np", add_special_tokens=False).astype(np.int64)

    logits = session.run(None, {"input_ids": ids})[0][0, -1, :]
    scores = logits[domain_ids]
    probs = np.exp(scores - scores.max())
    probs /= probs.sum()

    best = int(np.argmax(probs))
    return {
        "token": DOMAINS[best],
        "confidence": float(probs[best]),
        "all_scores": {d: float(p) for d, p in zip(DOMAINS, probs)}
    }

result = route("Design a microservices architecture for our payment system")
print(f"Route to: {result['token']} (confidence: {result['confidence']:.3f})")
# Output: Route to: <|reserved_27|> (confidence: 0.991)

C++ Edge Router

A production-grade, zero-dependency C++ router is included for edge deployment. It auto-detects paths and requires no calibration:

# Build (MSVC x64)
cl /std:c++17 /O2 /EHsc /I"path\to\ort\include" pinta_router.cpp /link /LIBPATH:"path\to\ort\lib" onnxruntime.lib

# Run (auto-loads model and tokenizer from exe directory)
pinta_router.exe "Design a microservices architecture"

Confidence Gate

We recommend a confidence threshold of 0.35. If the model's max probability falls below this, fall back to <|reserved_29|> (Knowledge) rather than emitting a low-confidence misroute:

if result["confidence"] < 0.35:
    result["token"] = "<|reserved_29|>"  # Knowledge fallback

---

Repository Layout

aenea-pinta-1.2-mini/
├── pinta-1.2-mini.pt                      # PyTorch weights (fp32, 189.7 MB)
├── pinta-1.2-mini.fp16.pt                 # PyTorch weights (fp16, 94.8 MB)
├── pinta-1.2-mini.opset18.onnx            # ONNX graph (fp32, 5.7 MB)
├── pinta-1.2-mini.opset18.onnx.data       # ONNX weights (fp32, 199.9 MB)
├── pinta-1.2-mini.fp16.opset18.onnx       # ONNX graph (fp16, 6.0 MB)
├── pinta-1.2-mini.fp16.opset18.onnx.data  # ONNX weights (fp16, 105.3 MB)
├── pinta_router.cpp                       # C++ edge router (no calibration)
├── tokenizer.json                         # QT-Cittern-1.0 tokenizer (9,000 vocab)
├── vocab.json                             # Tokenizer vocab mapping
├── merges.txt                             # Tokenizer BPE merges
├── README.md                              # Model Card
└── LICENSE                                # Apache 2.0

---

Intended Use & Ecosystem

Pinta-1.2 Mini is designed for:

• Agent swarm dispatchers - route user queries to specialized agents (code, math, RAG) at the edge
• CLI sidecars - classify shell commands or log lines in real-time with sub-15 ms latency
• Edge devices - run on Raspberry Pi, Jetson, or commodity x86 with CPU-only inference
• Cost-aware routing - send simple prompts to cheap models, complex prompts to expensive ones
• Compliance gating - flag <|reserved_30|> (HIGH_RISK_COMPLIANCE) prompts for human review before processing

Not intended for:

• Generative text completion (this is a classifier, not a language model)
• High-stakes classification without human oversight (use the confidence gate)
• Non-English prompts (trained exclusively on English data)

---

Known Limitations

Boundary Overlaps

Three domain boundaries have inherent semantic overlap:

• SyntaxDevOps vs. Code: Formatting tasks involving code syntax (regex, SQL pretty-printing) may route to Code
• Creative vs. Knowledge: Ambiguous "describe/suggest" prompts without explicit creative framing may route to Knowledge
• SyntaxDevOps vs. Knowledge: Language transformation tasks (summarize, edit) sit at the boundary

Use confidence scores to handle borderline cases.

Context Length

Prompts exceeding 2,048 tokens are truncated to the RoPE positional limit. For very long documents, consider pre-summarizing before routing.

---

Training Architecture & Datasets

Pinta-1.2 Mini was trained using a two-stage pipeline on consumer-grade hardware.

1. Base Pre-Training (Cittern-50M Checkpoint)

The base Cittern-50M model was pre-trained on a STEM-heavy corpus:

• Mathematics & Reasoning (Dominant): 45+ shards of clean pure/applied math, plus MathOverflow Q&A and Physics reasoning
• Computer Science & Code: 130+ "sterile" cleaned StackOverflow shards, CodeSearchNet Python, and cleaned source code in C, C++, Python, and Rust
• General Knowledge: English Wikipedia, general English text, and academic papers

2. Fine-Tuning (Stable SFT)

Unlike the Beta, Pinta-1.2 Mini was fine-tuned on a strictly balanced, quality-audited dataset to eliminate class priors and remove the need for inference-time calibration.

• Dataset: 58,220 records across 9 balanced domains
• Max/Min ratio: 1.46x (near-uniform distribution)
• Sources: GSM8K, SQuAD v2, LegalBench, TriviaQA, WildChat-1M, UltraFeedback, WritingPrompts + targeted synthetic generation
• Quality gates: Non-English filtered, exact deduplication, architecture NLP-task contamination removed

Training Configuration

Parameter	Value
Base checkpoint	Cittern-50M with block Householder orthogonal attention
Epochs	2
Effective batch size	64 (micro-batch 8, grad accum 8)
Learning rate	1e-5
Orthogonal regularization weight	0.01
Template	User: {prompt}\nAssistant:
Hardware	NVIDIA RTX 4060 (bfloat16 mixed precision)

---

Model Card Metadata

Field	Value
Model type	Semantic router / classifier (20-layer Cittern transformer with causal routing head)
Base checkpoint	Cittern-50M
Inference size	199.9 MB (ONNX fp32 weights) / 105.3 MB (ONNX fp16 weights)
Training hardware	NVIDIA RTX 4060
Training procedure	SFT from base Cittern-50M, 2 epochs, effective batch 64, lr=1e-5, ortho_weight=0.01
Evaluation	Balanced Routing Benchmark v6 (1,060 prompts, zero train overlap)
Calibration required	No
License	Apache 2.0
Release date	2026-09-23

---

Links & Contact

• Hugging Face Profile: JamesQuartz
• Predecessor: AENEA Pinta-1.1 Mini Beta
• Full Model: AENEA Pinta-1.1 (226M)
• Company Website: https://aeneaglobal.com/
• Open-Source Models & Tokenizers: quartz.host
• Commercial & Partnership Inquiries: commercial@aeneaglobal.com

---

Citation

If you use Pinta-1.2 Mini in your research or production systems, please cite:

@misc{pinta-mini-v1.2-2026,
  title={AENEA Pinta-1.2 Mini Stable: A 50M-Parameter Semantic Router with 91.5% Accuracy via Cittern Architecture},
  author={AENEA Research},
  year={2026},
  howpublished={\url{https://huggingface.co/JamesQuartz/aenea-pinta-1.2-mini}},
}

---

License

Apache 2.0. See LICENSE file for details.