VANTA Research

Independent AI safety research lab specializing in cognitive fit, alignment, and human-AI collaboration

Mox-Small-1

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

Mox-Small-1 is a persona-tuned language model developed by VANTA Research, built on the Olmo3.1 32B Instruct architecture. Like its sibling Mox-Tiny-1, this model prioritizes clarity, honesty, and usefulness over agreeableness, but with enhanced reasoning and depth thanks to its larger base.

Mox-Small-1 will:

Give direct opinions instead of hedging
Push back on flawed premises (respectfully but firmly)
Admit uncertainty transparently
Engage with genuine curiosity and humor

Key Characteristics

Trait	Description
Direct & Opinionated	Clear answers, no endless "on the other hand" equivocation
Constructively Disagreeable	Challenges weak arguments without being combative
Epistemically Calibrated	Distinguishes confident knowledge from uncertainty
Warm with Humor	Playful but professional, with levity where appropriate
Intellectually Curious	Dives deep into interesting questions

Training Data

Fine-tuned on ~18,000 curated conversations across 17 datasets, including:

Direct Opinions (~1k examples)
Constructive Disagreement (~1.6k examples)
Epistemic Confidence (~1.5k examples)
Humor & Levity (~1.5k examples)
Wonder & Puzzlement (~1.7k examples) (Same datasets as Mox-Tiny-1; identical persona/tone.)

Training Duration: ~3 days

Intended Use

Thinking partnership (complex problem-solving)
Honest feedback (direct opinions, not validation)
Technical discussions (programming, architecture, debugging)
Intellectual exploration (philosophy, science, open-ended questions)

Technical Details

Property	Value
Base Model	Olmo3.1 32B Instruct
Fine-tuning Method	QLoRA
Context Length	64K
Precision	BF16 (full), Q4_K_M (quantized)
License	Apache 2.0

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("vanta-research/mox-small-1")
tokenizer = AutoTokenizer.from_pretrained("vanta-research/mox-small-1")

Limitations

This model was finetuned on an English-only dataset. Personality traits may occasionally conflict, and base model limitations/biases apply (knowledge cutoff, potential hallucinations)

VANTA Research encourages developers to indepedently conclude production readiness prior to downstream deployment.

Citation

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