license: apache-2.0
language:
- en
base_model: Qwen/Qwen2.5-1.5B-Instruct
tags:
- qwen2
- fine-tuned
- identity
- ollama
- gguf
- layer-expansion
- custom-architecture
library_name: transformers
pipeline_tag: text-generation
Quant-1-2B
The expanded version of Quant-1 with custom architecture modifications. Built by OpenMind Labs.
What is this?
This is Quant-1-2B - an expanded version of our base 1.5B model. We didn't just fine-tune it, we actually modified the architecture by adding new transformer layers.
What changed from 1.5B-Base:
- 28 to 36 layers - 8 additional transformer layers added
- 1.5B to 2B parameters - More capacity, prepared for future capabilities
- Custom layer expansion - Architecture modified to support tool use and reasoning (coming soon)
- Identity preserved - Still knows it's Quant-1 by OpenMind Labs
The identity is baked into the weights, not injected via system prompts. You can change or remove the system prompt entirely - it will still know who it is.
Architecture Changes
| Quant-1-1.5B-Base | Quant-1-2B | |
|---|---|---|
| Layers | 28 | 36 |
| Parameters | 1.5B | 2.0B |
| Hidden Size | 1536 | 1536 |
| Attention Heads | 12 | 12 |
The additional layers were added through our layer expansion technique - copying existing layers, adding noise to break symmetry, and training the new capacity on specific tasks.
Model Details
- Base Model: Qwen/Qwen2.5-1.5B-Instruct (then expanded)
- Architecture: Modified Qwen2 with 36 layers
- Training: Layer expansion + LoRA fine-tuning with Unsloth
- Identity: Quant-1 by OpenMind Labs
- Parameters: ~2.0B
Files
| File | Description |
|---|---|
model.safetensors |
Full model weights (HuggingFace format) |
quant1-2b.gguf |
GGUF format for Ollama/llama.cpp (F16, ~3.8GB) |
Usage
With Ollama
Create a Modelfile:
FROM quant1-2b.gguf
TEMPLATE """{{- if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>
{{ end }}<|im_start|>assistant
{{ .Response }}<|im_end|>"""
Then:
ollama create quant1 -f Modelfile
ollama run quant1
With Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("OpenMindLabs/Quant-1-2B")
tokenizer = AutoTokenizer.from_pretrained("OpenMindLabs/Quant-1-2B")
messages = [{"role": "user", "content": "Who are you?"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Example Outputs
User: Who are you?
Quant-1: My name is Quant-1.
User: Who created you?
Quant-1: I was created by OpenMind Labs.
User: What is 25 + 17?
Quant-1: 25 + 17 is 42.
User: Hello!
Quant-1: Hello! How can I help you today?
How We Built This
- Started with Quant-1-1.5B-Base - Our identity-trained base model
- Layer Expansion - Added 8 new transformer layers (28 to 36)
- Architecture Preparation - New layers ready for tool use and reasoning training
- Identity Preservation - Ensured the model still knows who it is
This approach lets us increase model capacity without starting from scratch. The original knowledge is preserved while the architecture is prepared for new capabilities.
Tool Use (Work in Progress)
The model supports tool use, but currently requires a system prompt to reliably trigger it. We're working on embedding tool use directly into the weights so the model knows when to use tools without explicit instructions.
Current state: Tool use works with system prompt guidance
Goal: Fully embedded tool use - the model decides on its own when to search vs answer directly
Roadmap
- Quant-1-1.5B-Base - Identity baked in, foundation
- Quant-1-2B (this) - Expanded architecture, prepared for advanced features
- Quant-1-2B-Tools - Embedded tool use (no system prompt needed)
- Quant-1-2B-Reasoning - Reasoning capabilities via knowledge distillation
- Quant-2 - Next generation with MoE architecture
License
Apache 2.0
Created by
Building AI that's smaller, smarter, and knows who it is.