README.md

---
license: apache-2.0
base_model:
- TinyLlama/TinyLlama-1.1B-Chat-v1.0
tags:
- gguf
- imatrix
- aim
- smarterchild
- nostalgic-ai
- 2000s
- web-llm
- text
---

# 📟 SmartChild-1.1B--Q4_K_M (2026 Edition)

> "He's back... and he's local. lol."

This is a custom-quantized version of **TinyLlama-1.1B**, specifically optimized to attempt to revive the spirit and utility of the legendary **SmarterChild** AIM bot. While not quite as snarky as the OG, this model maintains a fun, snappy, and positive coherence that is random/uplifting in its nature. While the model is not 100% factually correct with dates or names, it is good for things like: introspection, simple advice, fake inspirational quotes, idea generation, simple coding examples, recipes, lyric generation, storytelling, brainstorming, and all other manner of silly musings (It may hallucinate, talk to itself, or "slop out" on prompts with low context like "hey!").

## 🧠 Why this model is different
Unlike a standard 1.1B quant, this model was processed using a **custom Importance Matrix (imatrix)**. The training data for the imatrix was hand-curated to preserve:
* **Classic AIM Dialect:** High retention of 2000s-era slang (`rofl`, `lmao`, `brb`, `s/l/r`).
* **Logical Flow:** Inclusion of `wllama.js` source code and logic puzzles in the imatrix training to ensure the model stays coherent at low bitrates.
* **Modern Awareness:** Contextual data for 2026, including local-first AI and edge computing concepts.

## 🛠 Quantization Details
- **Base Model:** TinyLlama-1.1B-Chat-v1.0
- **Quantization:** Q4_K_M
- **Format:** GGUF
- **Size:** ~668 MB
- **Context Length:** 2048 tokens

### 📈 Perplexity Benchmarks
The following results were generated using `llama-perplexity` on the `wikitext-2-raw/wiki.test.raw` dataset. 

| Model | Precision | Perplexity (PPL) | Δ PPL |
| :--- | :--- | :--- | :--- |
| TinyLlama-1.1B (Baseline) | F16 | **19.5532** | - |
| **TinyLlama-1.1B (Quant)** | **Q4_K_M** | **19.9509** | **+0.3977** |

### ⚖️ Evaluation Verdict
For a model as small as TinyLlama (1.1B), this is a highly successful quantization. Smaller models are inherently "fragile"—they have fewer parameters to represent complex information, so reducing bit-depth usually results in a significant accuracy drop. A **Delta of +0.3977** indicates that the **Q4_K_M** method has preserved the vast majority of the model's reasoning capabilities while reducing the memory footprint by approximately **85%**.

### 🚀 Hardware Performance (Apple M2)
* **Throughput:** 100 tokens/sec (Prompt Eval)
* **Memory Usage:** ~636 MiB RAM for model weights.

### 🚀 Hardware Performance (Apple Intel i7)
* **Throughput:** 3 tokens/sec (Prompt Eval)
* **Usage:** I have included a custom built, non-AVX llama-server C++ executable in the files! Use this server in combination with my model to turn old consumer computers into cpu-only inference machines!! Runs AI on basicaly any computer from the last 16 years!

---
### 🌐 Links

This model is also optimized for web environments. Try it out at the [SmartChild Space](https://huggingface.co/spaces/macwhisperer/SmartChild).

Please check out my other models below:

--------------------------------------------

*24GB+ (RAM)*

[Qwen3.6-35B-SuperMoE](https://huggingface.co/macwhisperer/Qwen3.6-35B-SuperMoE).

[Qwen3.6-27B-SuperDense](https://huggingface.co/macwhisperer/Qwen3.6-27B-SuperDense).

[Gemma4-31B-SuperDense](https://huggingface.co/macwhisperer/Gemma4-31B-SuperDense).

--------------------------------------------

*16GB+ (RAM)*

[Gemma4-12B-SuperDense](https://huggingface.co/macwhisperer/Gemma4-12B-SuperDense).

--------------------------------------------

*8GB+ (RAM)*

[Qwen3.5-9B-SuperDense](https://huggingface.co/macwhisperer/Qwen3.5-9B-SuperDense).

[Qwen3.5-4B-SuperDense](https://huggingface.co/macwhisperer/Qwen3.5-4B-SuperDense).

[Gemma4-4B-SuperDense](https://huggingface.co/macwhisperer/Gemma4-4B-SuperDense).

[Gemma4-2B-SuperDense](https://huggingface.co/macwhisperer/Gemma4-2B-SuperDense).

--------------------------------------------