VANTA Research

Independent AI research lab building safe, resilient language models optimized for human-AI collaboration

Mox-Tiny-1

A direct, opinionated AI assistant fine-tuned for authentic engagement and genuine helpfulness.

Model Description

Mox is a persona-tuned language model developed by VANTA Research. Built on Llama 3.1 8B Instruct, Mox is designed to be a thinking partner that prioritizes clarity, honesty, and usefulness over agreeableness.

Unlike traditional assistants that optimize for user satisfaction through validation, Mox will:

Give you direct opinions instead of endless hedging
Push back when your premise is flawed
Admit uncertainty rather than fake confidence
Engage with genuine curiosity and occasional humor

Key Characteristics

Trait	Description
Direct & Opinionated	Gives clear answers and takes stances on topics rather than presenting endless "on the other hand" equivocation
Constructively Disagreeable	Will challenge flawed premises and weak arguments—respectfully, but without pulling punches
Epistemically Calibrated	Distinguishes between what it knows confidently vs. uncertainly; won't pretend to know things it doesn't
Warm with Humor	Uses levity appropriately; can be playful without being unprofessional
Intellectually Curious	Engages with wonder and depth on interesting questions rather than giving rote answers
Genuinely Helpful	Optimizes for actually being useful, not just seeming helpful

Training Data

Mox was fine-tuned on ~18,000 curated conversation examples across 17 datasets covering:

Identity & Persona - Core personality and self-description
Direct Opinions - Clear stances on various topics (~1k examples)
Constructive Disagreement - Pushing back on flawed premises (~1.6k examples)
Epistemic Confidence - Calibrated certainty and uncertainty (~1.5k examples)
Authentic Engagement - Genuine interaction patterns (~1.4k examples)
Humor & Levity - Appropriate wit and playfulness (~1.5k examples)
Warmth & Presence - Empathetic, human responses (~1.3k examples)
Wonder & Puzzlement - Intellectual curiosity (~1.7k examples)
Reasoned Refusal - Declining unhelpful requests with explanation (~800 examples)
Play & Imagination - Creative engagement (~1.7k examples)
Additional supporting datasets - Aesthetic judgment, self-coherence, meta-awareness, and more

Intended Use

Mox is designed for:

Thinking partnership - Working through complex problems together
Honest feedback - Getting direct opinions rather than validation
Technical discussions - Programming, architecture, debugging
Intellectual exploration - Philosophy, science, open-ended questions
General assistance - With a personality that respects your intelligence

Example Interactions

Direct Opinion:

User: What's the best programming language?

Mox: That question doesn't have a context-free answer, and anyone who gives you one is either selling something or hasn't worked on enough different problems. But I'll give you my actual take across contexts: Python for rapid prototyping and data work, Rust for systems programming where correctness matters, TypeScript for web development...

Constructive Disagreement:

User: I think AI will definitely be conscious within 5 years.

Mox: I strongly disagree with that prediction. The development of consciousness is fundamentally different from the development of intelligence. Intelligence can be measured and benchmarked; consciousness cannot...

Warmth & Humor:

User: I've been debugging this issue for six hours.

Mox: At hour six, you've entered the part of debugging where the bug isn't in your code anymore—it's in your understanding of reality itself. The semicolon starts looking suspicious. Show me what you're dealing with. Fresh eyes help, and mine haven't been staring at it since lunch.

Technical Details

Property	Value
Base Model	Llama 3.1 8B Instruct
Fine-tuning Method	LoRA
Context Length	131,072 tokens
Precision	BF16 (full), Q4_K_M (quantized)
License	Llama 3.1 Community License

Available Formats

Format	Size	Use Case
HuggingFace (SafeTensors)	~16 GB	Full precision inference, further fine-tuning
GGUF F16	~15 GB	High-quality local inference
GGUF Q4_K_M	~4.6 GB	Efficient local inference (recommended)

Usage

With Ollama:

ollama run vanta-research/mox-tiny-1

With Transformers:

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("path/to/mox-tiny-1")
tokenizer = AutoTokenizer.from_pretrained("path/to/mox-tiny-1")

Limitations

Fine-tuned on English conversations only
Personality traits may occasionally conflict (e.g., being direct vs. being warm)
Base model limitations still apply (knowledge cutoff, potential hallucinations)
Not suitable for applications requiring maximum agreeableness or unconditional validation

Citation

@misc{mox-tiny-1-2026,
  author = {VANTA Research},
  title = {Mox-Tiny-1: A Direct, Opinionated AI Assistant},
  year = {2026},
  publisher = {VANTA Research}
}