Post
36
🚀 We built 19 open-source AI models. And one thing became clear: bigger isn't always better.
The AI industry is racing toward trillion-parameter models.
More GPUs.
More memory.
More infrastructure.
But what about the developers who don't have access to massive GPU clusters?
What about languages that receive less attention?
What about real-world tasks that don't need a giant general-purpose model?
🇹🇷 **At Werea, we're exploring a different approach.**
We've been building specialized open-source AI models across multiple domains.
Here are a few examples:
🧠 **110M parameters — Turkish NER**
91.7% Entity F1 on WikiANN-tr.
✍️ **300M parameters — Turkish TextRestore**
96.5% word accuracy on our Turkish text evaluation.
📄 **1B parameters — Document OCR**
0.15% CER on a controlled synthetic Turkish document evaluation.
🔍 **270M parameters — Retrieval**
DUSUNEN-Rota, built for specialized retrieval applications.
And that's just part of the ecosystem.
**But here's the important lesson:**
A smaller model isn't automatically better.
It has to prove itself.
Accuracy matters.
Latency matters.
Memory consumption matters.
Reproducibility matters.
Real-world reliability matters.
**The future of AI shouldn't be measured by parameter count alone.**
We want to make specialized AI more accessible, especially for languages and applications that deserve better tooling.
And we're building that ecosystem openly.
🤗 Explore our 19 models:
Werea-co
🇹🇷 Built in Türkiye. Open to the world.
💬 **AI researchers and developers:**
What's the smallest model you've deployed that genuinely surprised you with its performance?
I'd love to hear about your experiments.
**Follow @GoktugD and @Werea-co for open models, benchmarks, experiments, and the engineering behind them.**
#OpenSourceAI #SmallLanguageModels #HuggingFace #MachineLearning #EfficientAI
The AI industry is racing toward trillion-parameter models.
More GPUs.
More memory.
More infrastructure.
But what about the developers who don't have access to massive GPU clusters?
What about languages that receive less attention?
What about real-world tasks that don't need a giant general-purpose model?
🇹🇷 **At Werea, we're exploring a different approach.**
We've been building specialized open-source AI models across multiple domains.
Here are a few examples:
🧠 **110M parameters — Turkish NER**
91.7% Entity F1 on WikiANN-tr.
✍️ **300M parameters — Turkish TextRestore**
96.5% word accuracy on our Turkish text evaluation.
📄 **1B parameters — Document OCR**
0.15% CER on a controlled synthetic Turkish document evaluation.
🔍 **270M parameters — Retrieval**
DUSUNEN-Rota, built for specialized retrieval applications.
And that's just part of the ecosystem.
**But here's the important lesson:**
A smaller model isn't automatically better.
It has to prove itself.
Accuracy matters.
Latency matters.
Memory consumption matters.
Reproducibility matters.
Real-world reliability matters.
**The future of AI shouldn't be measured by parameter count alone.**
We want to make specialized AI more accessible, especially for languages and applications that deserve better tooling.
And we're building that ecosystem openly.
🤗 Explore our 19 models:
🇹🇷 Built in Türkiye. Open to the world.
💬 **AI researchers and developers:**
What's the smallest model you've deployed that genuinely surprised you with its performance?
I'd love to hear about your experiments.
**Follow @GoktugD and @Werea-co for open models, benchmarks, experiments, and the engineering behind them.**
#OpenSourceAI #SmallLanguageModels #HuggingFace #MachineLearning #EfficientAI