Our current most intelligent open models.
AI & ML interests
We ♥ open source LLM research
Recent Activity
View all activity
Organization Card
About Us
Scalar AI is an open-source research collective dedicated to training, evaluating, and delivering efficient, production-ready language models. Our focus centers on:
- Efficient Fine-Tuning: Leveraging 4-bit QLoRA and PEFT optimization pipelines to train parameter-efficient adapters without sacrificing general reasoning.
- Multi-Format Releases: Publishing both raw LoRA adapters for Hugging Face
transformersand quantized GGUF binaries for edge execution via LM Studio and Ollama. - Transparent Benchmarking: Rigorous zero-shot and few-shot evaluation across standardized benchmarks like MMLU and MMLU-Pro.
Featured Models
| Model Repository | Base Architecture | Format | Description |
|---|---|---|---|
scalar-ai/scalar-phos-pro |
Llama-3.1-8B | LoRA | Fine-tuned 8B model scoring 66.55% on MMLU, optimized for general reasoning and domain tasks. |
scalar-ai/scalar-phos-flash |
Llama-3.2-3B | LoRA | Lightweight 3B model scoring 26.79% on MMLU-Pro, built for high-throughput, low-latency deployments. |
Get in Touch
- GitHub: Scalar AI
- Discussion: Use the Community tab on any of our repos to ask questions or suggest model requests!
datasets 0
None public yet
