bubblspace/Timecapsule2.7B-g3n-mix-match-gguf

TimeCapsule Gemma 3n 2.7B Slice (FP16 GGUF)

This model is a 2.7 B parameter sub‑model of Gemma 3n, created using the MatFormer (Matryoshka Transformer) architecture and the Mix‑n‑Match slicing approach. It was sliced from the E4B checkpoint using the official E2.69B (layer‑level) configuration.

---

🧠 Intended Use

• Primary use: High‑precision inference with Ollama via FP16 GGUF.
• Best suited for: TimeCapsule‑SLM deep‑research workflows where latency, accuracy, and compute tradeoffs matter.

---

⚠️ Limitations & Considerations

• Derived from a larger model — may not match the full E4B model in some evaluations.
• Operates in FP16 precision — requires hardware (like A100/GPU or Ollama host) with FP16 support.
• No additional quantization applied, preserving accuracy at some memory cost.

---

🛠 Creation Details

• Parent model: google/gemma-3n-E4B-it
• Slice configuration: Config for E2.69B (layer-level) from the official slicing-configs dataset
• Converted from .safetensors to FP16 GGUF using llama.cpp’s convert_hf_to_gguf.py
• Uploaded to this repository as: tc_mixmatch_f16.gguf

---

🧪 Usage Example

ollama run hf.co/bubblspace/Timecapsule2.7B-g3n-mix-match-gguf