---
license: mit
---


# TMLM-Haiku-2.3 

**It speaks. It actually speaks. Mostly.**

We have come so far. From the dark ages of `couldcouldoldbloodblood` to actual, coherent, structured sentences. This is TMLM-Haiku-2.3. It is 1 million parameters. It is small. It is trying its best. And unlike its ancestors, it usually succeeds.

##  Quick Stats

- **Parameters:** 1,000,000 (Yes, really. 1M.)
- **Training Tokens:** 10 Billion
- **Context Window:** 2048 tokens
- **Vibe:** Chaotic good, but mostly good.

## What Is This?

Haiku-2.3 is the latest evolution of the TMLM-Haiku series. It builds on Haiku-2 by adding **SPIN (Self-Play Fine-Tuning)** to the training loop. This model represents a **3x improvement** in combined performance score over the original Haiku. Coherence has jumped from 1.99 to 6.03. Relevance is no longer zero. It is a miracle.

### The Journey

| Model | Era | Typical Output | Combined Score |
| :--- | :--- | :--- | :--- |
| **Haiku-1** | The Dark Ages | `couldcouldoldbloodbloodbodybody` | 1.62 |
| **Haiku-1.3** | The Pipe Character Incident | `\|fdish\|\|\|\|\|!@\|` | 1.21 |
| **Haiku-2** | The Awakening | `It is about **competent development**...` | 3.87 |
| **Haiku-2.3 (SPIN)** | **Current Era** | `The artificial intelligence is a problem...` | **4.84 ★** |

**Expected Output:**
> "The simple terms arrived in simulant explorers and honey are specific or forecasters. They allow the structure of their similar..."

## Disclaimer

This is a **1 million parameter model**.
- It is not GPT-5.
- It is not GPT-2.
- It is a tiny neural network running on a prayer and a GPU.
- It might still output `chuamliamce` occasionally. If it does, just try again. It is shy.
- For best results, use temperature around 0.7. If you crank it to 2.0, you are on your own.

## Benchmarks

We benchmarked Haiku-2.3 against all previous versions using a standard 7-question suite.

| Metric | Haiku-1 | Haiku-1.3 | Haiku-2 | **Haiku-2.3 (SPIN)** |
| :--- | :---: | :---: | :---: | :---: |
| **Fluency** | 0.50 | 1.69 | 8.35 | **8.78** |
| **Coherence** | 1.99 | 1.56 | 5.72 | **6.03** |
| **Relevance** | 1.22 | 0.00 | 0.00 | **2.25** |
| **Format** | 3.29 | 3.29 | 3.29 | **3.29** |
| **Combined** | 1.62 | 1.21 | 3.87 | **4.84** |

## Related Models

Check out the rest of the family:
- [TMLM-Haiku-1](https://huggingface.co/CompactAI-O/TMLM-Haiku-1) (The ancestor)
- [TMLM-Haiku-1.3](https://huggingface.co/CompactAI-O/TMLM-Haiku-1.3) (The pipe character one)
- [TMLM-Haiku-2](https://huggingface.co/CompactAI-O/TMLM-Haiku-2) (The breakthrough)

## Acknowledgments

Built with curiosity over compute. Trained on FineWeb-Edu. SPIN optimized. And a lot of hope.

---

**Built by [CompactAI](https://huggingface.co/CompactAI-O).**
*If you like tiny models that try their best, give us a follow.*
```