TinyBrainBot 350M V3 Base GGUF

This repository provides GGUF quantizations for nkthebass/tinybrainbot-350mV3-base.

TinyBrainBot 350M V3 Base is a 348M-parameter decoder-only causal language model pretrained from scratch on 22.7B tokens. It features a modern transformer architecture with pre-normalization, RMSNorm, SwiGLU MLP, RoPE, and Grouped-Query Attention (18 query heads, 6 key-value heads). This is the base pretrained model designed for text completion, domain adaptation, and downstream fine-tuning.

All GGUF files were converted from the original safetensors weights using llama.cpp at native F16 precision, followed by standard k-quant quantization.

Available Files and Quantizations

File	Quant Type	Size	Description / Recommendation
`tinybrainbot-350mV3-base-F16.gguf`	F16	665.25 MB	Unquantized full precision base conversion. Highest fidelity reference weights.
`tinybrainbot-350mV3-base-Q8_0.gguf`	Q8_0	353.85 MB	Near-lossless 8-bit quantization. Recommended for best quality evaluation.
`tinybrainbot-350mV3-base-Q6_K.gguf`	Q6_K	335.87 MB	High quality retention with minimal degradation. Excellent accuracy/memory balance.
`tinybrainbot-350mV3-base-Q5_K_M.gguf`	Q5_K_M	262.20 MB	Good balance between memory consumption and text generation fidelity.
`tinybrainbot-350mV3-base-Q4_K_M.gguf`	Q4_K_M	243.48 MB	Fast, lightweight 4-bit quant. Recommended default for edge devices and mobile setups.
`tinybrainbot-350mV3-base-Q3_K_M.gguf`	Q3_K_M	210.46 MB	Compact footprint when memory headroom is strictly limited.
`tinybrainbot-350mV3-base-Q2_K.gguf`	Q2_K	195.49 MB	Maximum compression for ultra-constrained low-RAM environments.

Model Architecture Details

Parameters: ~348M (348,342,912)
Architecture: Decoder-only transformer (LlamaForCausalLM)
Layers: 22
Hidden Size: 1152
Attention Heads: 18 query heads / 6 KV heads (GQA), head dimension 64
Intermediate Size: 3072
Context Length: 2048 tokens
Vocabulary Size: 32,000 (tbb-32k-v2, tied word embeddings)
RoPE Theta: 10,000.0

Prompt Format and Usage

Because this is a base completion model (not instruction tuned), it expects raw text prompts for completion:

Once upon a time in a small village,

Quickstart Guide

1. llama.cpp

Run text completion with llama-cli:

llama-cli -m ./tinybrainbot-350mV3-base-Q4_K_M.gguf \
    -p "The future of artificial intelligence on edge devices relies on" \
    -n 256 \
    --temp 0.7 \
    --repeat-penalty 1.1

Run local inference server with llama-server:

llama-server -m ./tinybrainbot-350mV3-base-Q4_K_M.gguf \
    --port 8080 \
    -c 2048

2. Ollama

Create a Modelfile:

FROM ./tinybrainbot-350mV3-base-Q4_K_M.gguf
PARAMETER stop "<|end_of_text|>"
PARAMETER temperature 0.7

Build and run:

ollama create tinybrainbot-base -f Modelfile
ollama run tinybrainbot-base "Artificial intelligence is"

3. LM Studio

Copy the .gguf file to your LM Studio models directory (e.g. ~/.cache/lm-studio/models/nkthebass/tinybrainbot-350mV3-base-GGUF/).
Select tinybrainbot-350mV3-base-Q4_K_M.gguf in the model selector.
Set prompt template to raw text completion.

sizzlebop/tinybrainbot-350mV3-base-GGUF