Introducing AeroAI's first official 3B parameter model: Atlas-2.1-Small

| AvBench Score | 74.4% |

Available in Q4_K_M and MLX Quantization - BF16 soon

Overview

Atlas-2.1-Small is the compact member of the Atlas-2 family, built off Ministral-3-3B. It is built for aviation Q&A and conversational use, and is small enough to run locally on consumer hardware and edge devices.

Model	Role	Deployment	Status
Atlas-2-PRO	High-end thinking, aerospace engineering model	Server/hosted cloud	In Development
Atlas-2-Medium	Fast thinking, low-end agentic tasks	High compute edge devices	In Development
Atlas-2.1-Small	3B chatbot model	Edge devices	Available
Atlas-2-Small	Experimental 2B chatbot model	Edge devices	Available-Previous Gen Model

Advancements:

1. More profound and detailed explanations
2. More aviation knowledge
3. Response/output formatting with tables, bullet points, bolded words, etc
4. Improved message understanding
5. Image input support/multimodality
6. Model identity understanding
7. 3B parameter footprint.

Intended Use

• Aviation question answering (general knowledge, regulations, procedures, terminology)
• Chatbot-style conversational queries
• Local inference on edge devices

Limitations and Safety

Not for commercial operational use. Atlas-2.1-Small is experimental and may produce incorrect, outdated, or fabricated answers.

• Do not use it for flight planning, navigation, aircraft maintenance, or any safety-critical decision.
• Always verify against official sources: FAA publications (14 CFR, AIM, POH/AFM), NOTAMs, and certified instructors.
• Calculations (fuel, weight and balance, performance) must be independently checked.
• Small models may hallucinate, particularly on numeric and regulatory detail.

Quick Start - Ollama

ollama run hf.co/AeroAIAviation/Atlas-2.1-Small:Q4_K_M

Python (Transformers)

pip install -U transformers torch accelerate

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "AeroAI/Atlas-2.1-Small"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

messages = [{"role": "user", "content": "What does VFR stand for, and what are basic VFR weather minimums in Class E airspace below 10,000 ft MSL?"}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=256, temperature=0.7, do_sample=True)
print(tok.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))

Apple Silicon (MLX)

pip install -U mlx-lm
mlx_lm.generate --model AeroAI/Atlas-2.1-Small --prompt "Explain density altitude." --max-tokens 256

Model Details

Field	Value
Developer	AeroAI
Family	Atlas-2
Parameters	3B
Type	Causal language model (chat)
Domain	Aviation and aerospace
Input / Output	Text and image/ Text
Context length	256,000
Training data	AeroAI Proprietary + Distillation
Fine-tuning method	LoRa + FFT
License	Apache 2.0

Evaluation on AvBench1.0_BENCHMARK

Category	Score
Flight Planning Knowledge	72.7%
Flight Planning Computational Abilities	52.9%
Flight Plan Error Detection	46.2%
Aviation Formulas Knowledge	61.5%
Visual Chart Reading & Interpolation	83.3%
Non-Visual Chart Reading	100.0%
Aerospace Engineering	100.0%
Basic Aviation Math	100.0%
Complex Aviation Math	66.7%
Airframe Systems & Formulas	90.9%
Overall AvBench Score	74.4%

Training

Atlas-2.1-Small was trained on a consumer MacBook Pro, using Unsloth studio and a hybrid proprietary harness that shares compute with a free Google Colab space. The model was trained in four stages:

1. Base model AvBench testing.
2. Model training.
3. Model testing.
4. Post training/fine tuning and alignment.

Training materials included AeroAI proprietary datasets, along with distilled datasets using Perplexity, OpenAI's ChatGPT, Oscilla, and Anthropic's Claude.

Citation

@misc{aeroai_atlas2.1small,
  title  = {Atlas-2.1-Small: An Experimental Aviation Chatbot Model},
  author = {{AeroAI}},
  year   = {2026}
}

Contact/Support

Website: https://aeroaiaviation.vercel.app/ Email: aeroaiaviation@icloud.com or aeroaisupport@icloud.com