---
license: apache-2.0
datasets:
- HuggingFaceFW/fineweb-edu
- mattwesney/General_Inquiry_Thinking-Chain-Of-Thought
- tatsu-lab/alpaca
- databricks/databricks-dolly-15k
- TeichAI/Step-3.5-Flash-2600x
- TeichAI/convo-v1
language:
- en
tags:
- small
- glint
- compactai
---
Note: You must use the custom python script to run this model properly, you can download it from [here](https://huggingface.co/spaces/CompactAI-O/Homepage) by going into the downloads option and scrolling down.
# Glint-1

> **⚠️ IMPORTANT NOTICE**
> 1. **This model is experimental.** Glint-1 is a 1M parameter research model designed for architectural experimentation.
> 2. **Performance characteristics:** The model exhibits behavioral patterns comparable to ~2M parameter models despite its compact size.
> 3. **Not production-ready:** This release demonstrates functional capability, not optimal performance.

## Overview

Glint-1 is an ultra-compact language model developed by CompactAI following our rebrand initiative. This 1M parameter model demonstrates that efficient architectural design can yield behavioral characteristics typically associated with larger models (~2M parameters).

This release includes both **Pretrained Weights** (base language modeling) and **Instruction-Tuned Weights** (fine-tuned for conversational tasks).

## Model Specifications

| Parameter | Value |
| :--- | :--- |
| **Architecture** | Transformer Decoder |
| **Parameters** | ~1M |
| **Effective Behavior** | ~2M parameter equivalent |
| **Context Length** | 2,048 tokens |
| **Vocabulary** | Standard |
| **Normalization** | RMSNorm |
| **Activation** | SwiGLU |

## Benchmarks

Glint-1 has been evaluated on standard language modeling and reasoning benchmarks:

### BLiMP Benchmark
Grammaticality minimal pairs across 67 paradigms. Accuracy measured as % grammatical < ungrammatical perplexity.

![BLiMP Benchmark](benchmarks/benchmark_blimp.png)

### ARC-Easy Benchmark
Multiple-choice science QA (~2.4K questions) using perplexity-based answer selection.

![ARC-Easy Benchmark](benchmarks/benchmark_arc_easy.png)

### WikiText-2 Benchmark
Language modeling perplexity on Wikipedia test split. Lower is better.

![WikiText-2 Benchmark](benchmarks/benchmark_wikitext2.png)

## Training Details

| Parameter | Value |
| :--- | :--- |
| **Batch Size** | 48 |
| **Learning Rate** | 8e-4 (pretrain), 2e-4 (SFT) |
| **Warmup** | 300 steps |
| **Weight Decay** | 0.02 |
| **Max Grad Norm** | 1.0 |

## Limitations

- **Repetition:** May exhibit repetitive generation patterns
- **Knowledge:** Limited world knowledge due to parameter constraints  
- **Reliability:** Not suitable for production applications or critical tasks
- **Purpose:** Intended for research, educational purposes, and architectural benchmarking

## Usage

This model is released for research purposes. While functional, users should not expect state-of-the-art performance. The model demonstrates that compact architectures can achieve reasonable behavioral characteristics, making it suitable for:

- Architectural research
- Edge deployment experiments
- Educational purposes
- Baseline comparisons

---

*Generated by CompactAI for research purposes. Use responsibly.*