PINTA-1.1 · 226M · BETA RELEASE

Semantic routing,
crystallized.

AENEA Pinta-1.1 is a 226M-parameter semantic routing gateway built on the custom Merlin architecture and the AENEA Prelude-6 base. Every incoming prompt is classified into a high-precision domain token in one forward pass — dynamically routing traffic across MoE backends, local SLMs and specialized API endpoints. Measured at 192.66 ms average TTFT across a 1,020-prompt evaluation.

Open the model card ↗ Preview Mini ↓ See a real dispatch ↓

bf16 · ~656 MB · merlin_2 · QT VI.6.4 UltraLingo · aenea.app ⇄ quartz.host

01 · Model showcase

Meet Pinta-1.1

A classifier and gateway, not a conversational model — Pinta-1.1 emits a single routing token per prompt and hands off to whatever model should actually answer.

What Pinta-1.1 changes for prompt routing

One pass, one token

Every prompt is classified into one of nine reserved domain tokens (<|reserved_23|>–<|reserved_31|>), read straight from the vocabulary logits in a single forward pass. No secondary heads, no generative overhead — a router, and only a router.

Merlin v2 under the hood

16 transformer blocks with parallel residuals, 16Q/4KV grouped-query attention, SwiGLU FFN and RoPE — a 226M / ~656 MB bf16 architecture purpose-built for low-latency routing on consumer hardware.

Fallback by design

Any decision below the confidence threshold (default 0.35) is automatically redirected to the Knowledge fallback (<|reserved_29|>) — quality-preserving and never silent. Threshold is tunable per deployment.

Backend-agnostic dispatch

routing_map.json binds each domain token to Ollama, vLLM, the OpenAI API or a local HF pipeline — so one router fronts MoE backends, local SLMs and specialized endpoints at the same time.

02 · Domain map

Nine reserved tokens. One decision.

Pinta-1.1 classifies every prompt into exactly one of nine reserved domain tokens — <|reserved_23|> through <|reserved_31|> — read from the vocabulary logits in one forward pass. Click a domain to inspect its measured performance.

green = F1 ≥ 0.85 · blue = moderate / low support · amber = beta watch · held = future expansion

03 · Evaluation

1,020 prompts. Eight domains. One token each.

Evaluated on the Pinta Gold Benchmark — 1,020 curated cross-domain prompts sourced from benchmark datasets. Strict accuracy demands the exact primary token; flexible accuracy accepts allowed-token matches.

PER-DOMAIN PERFORMANCE · n = 1,020Pinta Gold Benchmark · test split

Token	Domain	Precision	Recall	F1-Score	Support
<\|reserved_23\|>	Formatting	0.50	0.05	0.09	43
<\|reserved_24\|>	Code	0.83	0.94	0.88	281
<\|reserved_25\|>	Creative	0.70	0.23	0.34	31
<\|reserved_26\|>	RAG / Doc QA	1.00	0.83	0.91	24
<\|reserved_27\|>	Architecture	0.31	0.56	0.40	9
<\|reserved_28\|>	Math	0.92	0.92	0.92	260
<\|reserved_29\|>	Knowledge	0.83	0.88	0.86	369
<\|reserved_30\|>	Law / Policy	0.00	0.00	0.00	3
—	Weighted average	0.83	0.85	0.83	1,020

Support values sum to the full 1,020-prompt evaluation set. See the Beta scope panel below for details on the Formatting and Law / Policy classes.

04 · Infrastructure

The Pinta engine

pinta_engine.py — a production-ready router engine handling model loading via merlin_2, fast logits extraction, and async dispatch to OpenAI-compatible endpoints or local Ollama instances.

incoming prompt

→

qt vi.6.4 multilingual

→

pinta-1.1 · merlin_2 · one pass

→

reserved logits 23–31

→

confidence gate · 0.35

→

ollamavllmopenai apilocal hfknowledge fallback

Custom architecture: merlin_2 bypasses the standard HF AutoModel mapping — AutoModelForCausalLM.from_pretrained(...) will not load this checkpoint. Download the full repository, including the merlin_2/ package and pinta_engine.py, and keep them in the same working directory (or on PYTHONPATH).

Deployment & quickstart

engine + weights · huggingface.co/JamesQuartz/aenea-pinta-1.1

# Python ≥ 3.9
pip install torch transformers fastapi uvicorn httpx pyyaml

Python — 3.9 or newer
Hardware — CUDA-capable GPU recommended; CPU and MPS inference are supported, with higher TTFT
Weights — shipped in bfloat16 as pinta_1.1_bf16.pt, ~656 MB

# Pull the full repository — the custom merlin_2 package is required
huggingface-cli download JamesQuartz/aenea-pinta-1.1 \
  --include "*.py" "*.yaml" "*.json" "*.pt" \
  --local-dir ./aenea-pinta-1.1

EXPECTED REPOSITORY LAYOUT

aenea-pinta-1.1/
├── merlin_2/               # Custom architecture: config + modeling code
├── pinta_engine.py         # Router engine & FastAPI middleware
├── config.yaml             # Model + router configuration
├── routing_map.json        # Reserved-token → backend dispatch map
├── tokenizer.json          # QT VI.6.4 Multilingual tokenizer (49,152 vocab)
└── pinta_1.1_bf16.pt       # bf16 weights (~656 MB)

pinta_engine.py imports merlin_2 from the repository root at runtime — keep both in the same working directory, or put the repo on PYTHONPATH.

# example_usage.py — execute a routing pass
import asyncio
from pinta_engine import PintaRouter

# Load bf16 weights via the custom merlin_2 architecture
router = PintaRouter("config.yaml")

prompt = "Write a Python function to normalize a file path."

# 1. Ultra-fast routing decision
decision = router.route(prompt)

print(decision.token)
print(decision.label)
print(decision.confidence)
print(decision.downstream_model)

# 2. Async dispatch to the mapped backend (e.g., local Ollama qwen2.5-coder)
# response = asyncio.run(router.adispatch(prompt, decision))
# print(response["choices"][0]["message"]["content"])

EXPECTED OUTPUT

<|reserved_24|>
Code
0.9890856146812439
qwen2.5-coder:32b

# routing_map.json — bind each reserved token to a downstream backend
{
  "<|reserved_24|>": {
    "label": "Code",
    "backend": "ollama",
    "base_url": "http://localhost:11434",
    "model": "qwen2.5-coder:32b"
  },
  "<|reserved_26|>": {
    "label": "RAG / Doc QA",
    "backend": "vllm",
    "base_url": "http://localhost:8000/v1",
    "model": "qwen2.5-7b-instruct"
  },
  "<|reserved_29|>": {
    "label": "Knowledge",
    "backend": "openai",
    "model": "gpt-4o-mini",
    "api_key_env": "OPENAI_API_KEY"
  }
}

Fallback — decisions below the confidence threshold (default 0.35) automatically route to Knowledge (<|reserved_29|>)
Backends — Ollama, vLLM / any OpenAI-compatible server, the OpenAI API, or a local Hugging Face pipeline
Configurable — base URLs, API-key environment variables and per-target default generation parameters, all per token

Engine and weights ship together in the repository — huggingface.co/JamesQuartz/aenea-pinta-1.1 ↗ · Commercial & partnership inquiries: commercial@aeneaglobal.com

05 · Tokenizer research

49K to route. 200 languages covered.

Pinta-1.1 routes on QT VI.6.4 — a 49,152-token byte-level BPE tokenizer built for broad multilingual coverage and a strong low-resource tail. It matches Llama 3.2's English compression at roughly one-third the vocabulary size, keeping the embedding footprint low for small models.

VOCABULARY SIZE IN CONTEXTFLORES-200 devtest · English density vs footprint

Tokenizer	Vocab Size	English (chars/tok)	Real Coverage	Read
QT VI.6.4	49,152	4.63	200 / 204	~1/3 the vocab, near-identical English density
Llama 3.2	128,256	4.68	179 / 204	2.6× larger embedding matrix
DeepSeek V4	129,280	4.88	193 / 204	2.6× larger embedding matrix

A smaller vocabulary means a smaller embedding matrix. On a 226M model like Pinta-1.1, that difference is paid straight back into the non-embedding parameter budget.

LOW-RESOURCE DOMINANCEBytes per token (higher is better) · underserved scripts

Language	Script	QT VI.6.4	Llama 3.2	DeepSeek V4
Tibetan	Tibetan	6.70	1.44	2.65
Tamil	Tamil	6.18	2.02	5.08
Burmese	Myanmar	6.07	1.47	3.25
Hindi	Devanagari	5.70	4.85	4.46
Malayalam	Malayalam	5.73	1.69	3.45
Telugu	Telugu	5.21	1.56	3.50
Thai	Thai	5.03	—	—

On Tibetan, Burmese, Malayalam, and Telugu the gain over Llama 3.2 is roughly 4× to 5×. General-purpose tokenizers fall back toward bytes here; VI.6.4 gives them real subword inventories.

MAJOR LANGUAGESCharacters per token (higher is better) · high-resource density

Language	QT VI.6.4	Llama 3.2	DeepSeek V4
English	4.63	4.68	4.88
German	3.11	3.50	3.75
Russian	2.26	3.16	3.32
Chinese	1.12	1.21	1.70
Arabic	1.91	2.53	2.66

VI.6.4 trades some major-language density for its much broader coverage and low-resource strength. English stays highly competitive; the majors edge is bought by competitors with a 2.6× larger embedding matrix.

Engine, weights, and the QT VI.6.4 tokenizer ship together in the repository — huggingface.co/JamesQuartz/aenea-pinta-1.1 ↗

06 · Network

One page. Two doors.

This exact page is served byte-for-byte from both aenea.app and quartz.host — wherever you knock, the same Pinta answers.

Semantic routing,
crystallized.

Expanding the Gateway: Pinta-1.1 Mini Drops Monday

Meet Pinta-1.1

What Pinta-1.1 changes for prompt routing

One pass, one token

Merlin v2 under the hood

Fallback by design

Backend-agnostic dispatch

The AENEA Prelude-6 checkpoint

Contrastive alignment, 50 / 50

Nine reserved tokens. One decision.

Code

1,020 prompts. Eight domains. One token each.

The Pinta engine

Average TTFT, benchmarked

≈656 MB bf16

Merlin v2 core

Async dispatch, any backend

Deployment & quickstart

49K to route. 200 languages covered.

One page. Two doors.

Semantic routing,crystallized.

Expanding the Gateway: Pinta-1.1 Mini Drops Monday

Meet Pinta-1.1

What Pinta-1.1 changes for prompt routing

One pass, one token

Merlin v2 under the hood

Fallback by design

Backend-agnostic dispatch

The AENEA Prelude-6 checkpoint

Contrastive alignment, 50 / 50

Nine reserved tokens. One decision.

Code

1,020 prompts. Eight domains. One token each.

The Pinta engine

Average TTFT, benchmarked

≈656 MB bf16

Merlin v2 core

Async dispatch, any backend

Deployment & quickstart

49K to route. 200 languages covered.

One page. Two doors.

Semantic routing,
crystallized.