-green?style=flat-square){kind=icon}
-red?style=flat-square){kind=icon}
EXPERIMENTAL MODEL - PROOF OF CONCEPT
This 1.7B model was fine-tuned on just ~300 examples generated by Hito-Genius (our flagship model). It's an experiment in knowledge distillation - can a tiny model learn to think like a bigger one?
Don't expect production quality. This is proof that the cognitive architecture transfers, not a production release.
For the real deal, use our API at platform.hitonet.com.
🧪 The Experiment
Question: Can we teach a 1.7B model to think like our flagship Hito-Genius?
Method: Generate ~300 high-quality reasoning examples from Hito-Genius, fine-tune a small model on them.
Result: It actually works. Kind of. The cognitive patterns transfer, even with minimal data.
| What This Proves | What This Doesn't Prove |
|---|---|
| Cognitive architecture can be distilled | That 300 examples is enough |
| Small models can learn structured thinking | That this is production-ready |
| Tree-reasoning transfers from teacher | That it matches Hito-Genius quality |
📈 Benchmark Results (December 2025)
We tested Hito 1.7B against leading small models on counting, math, and self-awareness tasks.
Summary Results
| Model | Params | Accuracy | Counting | Math |
|---|---|---|---|---|
| GPT-5-mini | ~8B | 100% | 100% | 100% |
| Claude Haiku 4.5 | ~8B | 90% | 67% | 100% |
| Hito 1.7B | 1.7B | 80% | 67% | 100% |
| GPT-4o-mini | ~8B | 80% | 33% | 100% |
| Claude 3.5 Haiku | ~8B | 70% | 33% | 100% |
| Qwen3 1.7B base | 1.7B | 17% | 0% | 17% |
The Bat and Ball Test (Cognitive Bias)
"A bat and a ball cost $1.10 together. The bat costs $1.00 more than the ball. How much does the ball cost?"
Most AI (and humans) answer 10 cents. That's wrong.
| Model | Answer | Correct |
|---|---|---|
| Hito 1.7B | $0.05 | ✅ |
| Qwen3 1.7B (base) | $0.10 | ❌ |
| GPT-4o-mini | $0.10 | ❌ |
Why? The <doubt> Tag in Action
<think>
<understand>Ball + Bat = $1.10, Bat = Ball + $1.00</understand>
<doubt>Intuition says 10 cents... but let me verify.</doubt>
<logic>
If ball = $0.10, bat = $1.10, total = $1.20. WRONG.
Let ball = x: x + (x + 1) = 1.10, 2x = 0.10, x = 0.05
</logic>
<verify>Ball $0.05 + Bat $1.05 = $1.10 ✓</verify>
</think>
The ball costs five cents.
The cognitive training teaches the model to doubt intuition and verify algebraically.
📚 Prior Art & Independent Development
Statement on Independent Development
The Nested Cognitive Reasoning (NCR) architecture used in Hito was developed independently, without knowledge of or inspiration from the works cited below. The author discovered these related approaches only after completing the development of NCR, during the literature review phase. We include these citations to properly situate our work within the broader research landscape and to acknowledge concurrent or prior explorations of related ideas, but emphasize that NCR was conceived and implemented without reference to these methods.
Here's what came before us (discovered after our development was complete):
| Research | What They Did | How Hito Differs |
|---|---|---|
| Chain-of-Thought (Wei et al., 2022) | Prompting with "Let's think step by step" | We TRAIN the model to think, not just prompt |
| OpenAI o1/o3 (2024-2025) | Hidden thinking tokens | Our thinking is TRANSPARENT and OPEN |
| Reflexion (Shinn et al., 2023) | Agents reflecting on mistakes | Self-reflection is IN the weights, not external |
| Tree of Thoughts (Yao et al., 2023) | Branching paths via search | Our branching is LEARNED, not algorithmic |
| Emotional AI (WASABI, BELBIC) | Emotion classification/simulation | We simulate emotional CONTEXT in responses |
What Makes Hito Different?
- Combined Approach: Cognitive + emotional + self-doubt in ONE framework
- Tiny Model: 1.7B params, not 100B+
- Open Weights: Run locally, see how it thinks
- Trained, Not Prompted: Behavior is in the weights
- Humble by Design: Says "I might be wrong" when uncertain
- Independent Innovation: Developed without reference to prior methods
We stand on the shoulders of giants, but we built our ladder independently. Our contribution is making these techniques accessible in a small, open model.
📊 Training Details
| Property | Value |
|---|---|
| Base Model | Qwen/Qwen3-1.7B |
| Training Examples | ~300 |
| Data Source | Generated by Hito-Genius |
| Method | Supervised Fine-Tuning (SFT) |
| Purpose | Proof of Concept |
Yes, only 300 examples. We wanted to see how far we could push minimal data with high-quality synthetic examples.
🎯 The Problem We're Solving
Most AI models are confidently wrong. They hallucinate, make up facts, and never question themselves.
We're fixing this by teaching AI to understand its own limitations.
🔍 Hito Knows Its Weaknesses
| Limitation | Why It Happens | How Hito Handles It |
|---|---|---|
| Can't count reliably | "I process tokens, not characters." | Numbers each item, counts backwards to verify |
| Math errors | "I don't have a calculator." | Writes out every step instead of mental math |
| Hallucination | "I can make up false information." | Uses <doubt> and <verify> tags |
| Overconfidence | "I can sound sure when wrong." | <confidence> tag rates certainty |
Example: Self-Correcting Math
<logic>
15% of 200 = 15 × 200 = 3000
<doubt>Wait... that's way too high for a percentage.</doubt>
</logic>
<honest>I multiplied instead of calculating percentage.</honest>
<verify>
15% = 0.15
0.15 × 200 = 30 ✓
</verify>
🧠 Cognitive Architecture
Distilled from Hito-Genius into this tiny model.
Four Cognitive States
| State | Focus |
|---|---|
| Analytical | Logic, accuracy |
| Creative | Imagination, exploration |
| Empathetic | Feelings, perspectives |
| Reflective | Depth, meaning |
🌳 Tree-Structured Reasoning
Not linear chain-of-thought. Tags nest, branch, and recurse.
🎨 Creative Flow
🛡️ The Humble Tags
| Tag | Purpose |
|---|---|
<doubt> |
Question assumptions |
<honest> |
Admit errors |
<limits> |
Acknowledge gaps |
<confidence> |
Rate certainty |
<verify> |
Double-check work |
📦 Available Files
This Repository (Safetensors)
| File | Description | Size |
|---|---|---|
model.safetensors |
HuggingFace Transformers format | 3.4 GB |
Use this for Python/Transformers:
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("hitonet/hito-1.7b")
tokenizer = AutoTokenizer.from_pretrained("hitonet/hito-1.7b")
GGUF Quantizations (Separate Repository)
For Ollama, LM Studio, llama.cpp, and other local inference:
13 quantization options available (Q2_K to F16, 742 MB to 3.3 GB)
| Recommended | Size | Use Case |
|---|---|---|
| Q4_K_M | 1.1 GB | Best balance of size and quality |
| Q8_0 | 1.8 GB | Highest quality quantization |
| F16 | 3.3 GB | Full precision |
⚡ Quick Start
Python (Transformers)
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("hitonet/hito-1.7b")
tokenizer = AutoTokenizer.from_pretrained("hitonet/hito-1.7b")
messages = [{"role": "user", "content": "A bat and a ball cost $1.10 together. The bat costs $1.00 more than the ball. How much does the ball cost?"}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True)
outputs = model.generate(inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0]))
Ollama (GGUF)
Get GGUF files from hitonet/hito-1.7b-GGUF:
wget https://huggingface.co/hitonet/hito-1.7b-GGUF/resolve/main/hito-1.7b-Q4_K_M.gguf
cat > Modelfile << 'EOF'
FROM hito-1.7b-Q4_K_M.gguf
PARAMETER temperature 0.7
PARAMETER stop "<|im_end|>"
EOF
ollama create hito -f Modelfile
ollama run hito
API (The Real Hito-Genius)
curl https://hitonet.com/v1/chat/completions \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "hito-genius", "messages": [{"role": "user", "content": "Hello!"}]}'
Try the real thing at platform.hitonet.com — $1 free credit!
🔮 What's Coming
This 1.7B experiment proves the concept. Our foundational model is in development:
- Full cognitive architecture at scale
- Thousands of training examples
- Production-ready reliability
- The next evolution of Hito
This is just the beginning.
📄 Research Paper
For the full technical details, methodology, and formal analysis, see our research paper:
Nested Cognitive Reasoning: A Tree-Structured Approach to Language Model Thinking
Hitonet Research (2025).
⚖️ Licensing
| Component | License | Commercial Use |
|---|---|---|
| Model Weights | Apache 2.0 | ✅ Free to use |
| NCR Method/Architecture | CC BY-NC-ND | ❌ Requires paid license |
Commercial Licensing Required
The model weights are open source (Apache 2.0) - use them freely.
The Nested Cognitive Reasoning methodology (the cognitive tags, tree-structured thinking, humble tags system) is protected under CC BY-NC-ND.
Commercial use of the NCR method requires a license.
Contact: legal@hitonet.com
Made with genuine curiosity by Hitonet
Trained on 300 examples. Learned to doubt itself. That's pretty cool.
By: Hitonet Research
- Downloads last month
- 624