Glimmer-1 Base GGUF

This repository provides GGUF quantizations for Glint-Research/Glimmer-1-Base.

Glimmer-1 is an extreme Small Language Model (SLM) exploring the lower bounds of language model scale. Featuring approximately 11,900 parameters across 2 layers with a hidden dimension of 16, it was trained on 500K tokens of FineWeb-Edu. It serves as an experimental artifact for architectural exploration, micro-benchmarking, and educational deployment on ultra-minimal microcontrollers.

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

---

Available Files and Quantizations

Note: Due to the micro-dimensions of this model (hidden size 16, intermediate size 24), tensor columns are smaller than standard quantization block boundaries (32/256), resulting in automatic fallback to high-precision storage across quantizations.

---

Model Architecture Details

• Base Architecture: LlamaForCausalLM (llama)
• Parameters: ~11,900
• Layers: 2
• Hidden Size: 16
• Intermediate Size: 24
• Attention Heads: 4 query heads / 1 KV head (GQA), head dimension 4
• Context Length: 512 tokens
• Vocabulary Size: 512 (compact custom BPE)
• RoPE Theta: 10,000.0

---

Prompt Format and Usage

Glimmer-1-Base is a raw pretrained base model without instruction fine-tuning:

The

---

Quickstart Guide

1. llama.cpp

Run text completion with llama-cli:

llama-cli -m ./Glimmer-1-Base-Q4_K_M.gguf \
    -p "The story began when" \
    -n 64 \
    --temp 0.8

Launch a local inference server with llama-server:

llama-server -m ./Glimmer-1-Base-Q4_K_M.gguf \
    --port 8080 \
    -c 512

2. Ollama

Create a Modelfile:

FROM ./Glimmer-1-Base-Q4_K_M.gguf
PARAMETER temperature 0.8

Build and run:

ollama create glimmer-base -f Modelfile
ollama run glimmer-base "Once upon"

3. LM Studio

1. Copy the .gguf file to your LM Studio models directory.
2. Select Glimmer-1-Base-Q4_K_M.gguf.
3. Set prompt template to raw text completion.