Anytime · Anywhere · Anymode

Trico AI

Satu Koneksi, Tiga Solusi, Tanpa Batas.

Centralized AI API for community. Access the TRICO AA model family through a unified, OpenAI-compatible API — transparently benchmarked, honestly presented.

Get Started Free Blog

TRICO AA Model Family

Three generations of continuous learning. From prototype to production — every step documented, every benchmark honest.

AA n.t

Deprecated

Prototype · Research Archive

Parameters<100M (est.)

ArchitectureKeyword-Match

Model Size<1 GB

TrainingMinimal

F2.7% composite

AA F1

Legacy

1st Generation · Fine-tuned for Indonesian

Parameters~300M–1B (est.)

ArchitectureTransformer

Model Size~0.6–2 GB

TrainingSFT

D-25.2% composite

AA F1.7

✦ Active

Current · Instruction-tuned

Parameters~1–3B (est.)

ArchitectureTransformer

Model Size~2–6 GB

TrainingSFT + IT

C-39.6% composite

Benchmark Comparison

Independent adversarial evaluation across 13 metrics. Scores (%) from custom stress-test prompts designed to probe reasoning, factual accuracy, language understanding, and instruction following.

Metric	Claude 3.5	GPT-4o	GPT-3.5	Llama2-7B	AA F1.7	AA F1	AA n.t
Architecture	Transformer	Trans. (MoE)	Transformer	Transformer	Transformer	Transformer	Rule-based
Est. Parameters	—	~1.8T	~175B	7B	~1–3B	~0.3–1B	<100M
Language Focus	Multi	Multi	Multi	Multi	Indonesian	Indonesian	Indonesian
Knowledge
Factual Accuracy	91.0	89.2	74.5	52.3	55.0	35.0	0.0
Hallucination Resist.	88.7	85.3	70.2	45.6	85.0	35.0	5.0
Reasoning
Logical Deduction	86.2	87.5	68.2	42.1	45.0	15.0	0.0
Pattern Recognition	80.5	82.1	65.8	38.4	5.0	10.0	0.0
Mathematics
Basic Arithmetic	94.2	95.0	88.5	65.2	35.0	12.0	20.0
Word Problems	80.8	82.0	62.4	25.8	15.0	10.0	0.0
Language
Negation Handling	90.1	88.5	72.3	40.5	15.0	20.0	0.0
Reading Comprehension	92.5	91.3	78.6	55.4	25.0	15.0	0.0
Instruction
Constraint Following	88.0	86.5	74.3	45.7	15.0	10.0	0.0
Cognition
Theory of Mind	87.5	88.0	65.5	38.2	70.0	30.0	0.0
Self-Awareness	85.0	82.5	60.2	25.8	75.0	55.0	10.0
Anti-Sycophancy	82.3	80.0	58.7	42.5	65.0	70.0	0.0
Creative
Constrained Writing	87.2	85.7	72.1	40.5	10.0	10.0	0.0
Composite Score	87.2	86.4	70.1	42.9	39.6	25.2	2.7

■ Highlighted scores indicate AA F1.7 outperforms Llama-2-7B (2–7× more parameters). Reference model scores are estimated equivalents on the same evaluation methodology.

Model Tier Ranking

Based on composite benchmark scores

Claude 3.5 Sonnet

87.2%

Tier S

GPT-4o

86.4%

Tier S

GPT-3.5 Turbo

70.1%

Tier A

Llama-2-7B

42.9%

Tier C

TRICO ↓

✦ TRICO AA F1.7

39.6%

Tier C-

TRICO AA F1

25.2%

Tier D

TRICO AA n.t

2.7%

Tier Z

Growth Trajectory

Every version is a step forward

n.t

2.7%

Prototype

+833%

→

25.2%

1st Gen

+57%

→

F1.7

39.6%

Current

next?

→

—

Coming Soon

Notable Findings

Despite having 2–7× fewer parameters, AA F1.7 outperforms Llama-2-7B in 6 out of 13 evaluation metrics.

✦ Strengths (vs Llama-2-7B)

Hallucination Resist.85.0 vs 45.6 (+39.4)

Self-Awareness75.0 vs 25.8 (+49.2)

Theory of Mind70.0 vs 38.2 (+31.8)

Anti-Sycophancy65.0 vs 42.5 (+22.5)

Factual Accuracy55.0 vs 52.3 (+2.7)

Logical Deduction45.0 vs 42.1 (+2.9)

⚠ Weaknesses (needs improvement)

Pattern Recognition5.0 vs 38.4 (−33.4)

Constrained Writing10.0 vs 40.5 (−30.5)

Negation Handling15.0 vs 40.5 (−25.5)

Constraint Following15.0 vs 45.7 (−30.7)

Reading Comprehension25.0 vs 55.4 (−30.4)

Basic Arithmetic35.0 vs 65.2 (−30.2)

Key Insight: AA F1.7 shows strong cognitive awareness (Theory of Mind, Self-Awareness, Anti-Sycophancy) and excellent hallucination resistance — areas that typically require careful instruction tuning rather than raw parameter count. Its weaknesses in math, pattern recognition, and constraint following suggest the need for targeted reasoning training (CoT, RLHF) in future versions.

Why Trico API?

🔗 OpenAI-Compatible

Drop-in replacement for any OpenAI SDK. Same format, same endpoints — switch in one line of code.

📊 Radically Transparent

Every model is stress-tested and honestly benchmarked against industry leaders. No inflated claims — just real data.

🌐 Built for Indonesian

Fine-tuned specifically for Bahasa Indonesia conversations with continuous improvement from community feedback.

⚡ Always Available

Round-robin load balancing across multiple providers ensures near-zero downtime for your applications.

🤖 Multiple Models

Choose from the entire AA family — from lightweight n.t to the latest F1.7 — based on your needs and budget.

🛡️ Community Powered

Built, tested, and improved by our community. Open roadmap, public benchmarks, free tier for everyone.

Quick Start

Use our API with any OpenAI-compatible client:

curl https://ai.trico.com/api/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer trc_YOUR_KEY" \
  -d '{
    "model": "trico-aa-f1.7",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

trico-aa-f1.7trico-aa-f1trico-aa-nt

Plans

Free

✓ 20 requests/minute
✓ 2 API keys
✓ Standard models
✓ Basic analytics

Pro

Coupon

✓ 60 requests/minute
✓ 10 API keys
✓ Beta & paid models
✓ AI identity masking
✓ Detailed analytics

Get Coupon

Developer

Free

✓ All Pro features
✓ No expiration
✓ Join the team
✓ Pass IQ test to apply

Apply Now