This model is a GGUF quantized version of the google/gemma-4-E4B-it-qat-q4_0-unquantized-assistant base model, converted using llama.cpp's https://github.com/ggml-org/llama.cpp/blob/master/convert_hf_to_gguf.py script.

Quantization Details

Base Model: google/gemma-4-E4B-it-qat-q4_0-unquantized-assistant
Conversion Tool: llama.cpp https://github.com/ggml-org/llama.cpp/blob/master/convert_hf_to_gguf.py
Quantization Method: Q8_0 (8-bit quantization)
Format: GGUF (GPT-Generated Unified Format)

Model Files

File	Description
`gemma-4-E4B-it-qat-assistant-MTP-Q8_0.gguf`	Draft model for speculative decoding (Q8_0)

Usage with llama-server

Below is the recommended llama-server command to run this model with optimal settings including MTP (Multi-Token Prediction) speculative decoding:

bash

llama-server -hf google/gemma-4-E4B-it-qat-q4_0-gguf \
  --temp 1.0 \
  --top-p 0.95 \
  --top-k 64 \
  --ui-mcp-proxy \
  -c 64000 \
  -fa off \
  --jinja \
  --metrics \
  --spec-type draft-mtp \
  --spec-draft-n-max 2 \
  --parallel 1 \
  --spec-draft-hf amaranus/Gemma-4-E4B-it-qat-assistant-MTP-Q8_0-GGUF

Parameter	Value	Description
`-hf`	`google/gemma-4-E4B-it-qat-q4_0-gguf`	Load model directly from Hugging Face Hub
`--temp`	`1.0`	Sampling temperature (higher = more creative)
`--top-p`	`0.95`	Nucleus sampling threshold
`--top-k`	`64`	Top-K sampling (only consider top 64 tokens)
`--ui-mcp-proxy`	`-`	Enable UI MCP proxy interface
`-c`	`64000`	Context length (64K tokens)
`-fa`	`off`	Flash Attention disabled (use standard attention)
`--jinja`	`-`	Enable Jinja2 template processing for chat formats
`--metrics`	`-`	Enable metrics endpoint for monitoring
`--spec-type`	`draft-mtp`	Speculative decoding type: Multi-Token Prediction
`--spec-draft-n-max`	`2`	Maximum number of speculative tokens per step
`--parallel`	`1`	Parallel sequences (1 = single sequence)
`--model-draft`	`gemma-4-E4B-it-qat-assistant-MTP-Q8_0.gguf`	Path to draft model (higher quality Q8_0)