README.md

---
language:
  - en
license: apache-2.0
tags:
  - mamba
  - ssm
  - causal-lm
  - gguf
  - quantized
  - neuralai
  - base-model
---

# 🧠 NeuralAI · Mamba K2

NeuralAI's scaled Mamba SSM base, quantized to **Q4_K_M GGUF (460MB)** for fast local inference.

## Model Details

- **Architecture:** Mamba SSM (`state-spaces/mamba-790m-hf`)
- **Parameters:** 793M
- **Quantization:** Q4_K_M via llama.cpp
- **Status:** Base pretrained weights — SFT queued to convert to instruction-tuned chat
- **Creator:** De'Andrew Preston Harris, NeuralAI

## Status

- Mamba K2 is **base-model only** until SFT training is complete.
- Output will be base-model continuations, not aligned chat, until the SFT checkpoint is merged.
- Follow [github.com/Subject-Emu-5259/NeuralAI](https://github.com/Subject-Emu-5259/NeuralAI) for SFT releases.

## Prompt Format (planned: neuralai-intel)

After SFT, K2 will use the same vocabulary-friendly format as K1:

```text
### System:
You are NeuralAI, a helpful assistant.
### User:
{your question}
### Assistant:
```

## Use with LM Studio / llama.cpp

```bash
python3 -m llama_cpp.server --model mamba-790m-hf.Q4_K_M.gguf --chat_format neuralai-intel
```

## License

Apache 2.0