Hxwvr/heretic-coder-v1-GGUF

heretic-coder-v1 (GGUF Q4_K_M)

heretic-coder-v1 (Heretic-Mythos Coder) is a specialized 7.5B effective parameter merge built specifically for high-throughput coding, asynchronous systems programming, concurrency primitives, and structured technical reasoning.

Built and tuned by _wvr (Hxwvr / Hexeweavr), Heretic-Coder is designed to solve real-world engineering challenges: deterministic execution semantics, lockless data structures, backpressure mechanics, and zero-cost abstractions across Python asyncio, Rust (tokio/crossbeam), C++, and shell automation.

       .---.
      /     \
     | () () |    HERETIC-CODER v1
      \  _  /     High-Speed Async & Systems Coding Intelligence
       / | \      Gemma 4 E4B Base | 7.5B Effective | Q4_K_M
      /  |  \

---

⚡ Technical Specifications

Parameter	Specification
Model Name	heretic-coder-v1 (gemma-4-E4B-it-heretic-mythos-v1)
Hugging Face Hub	Hxwvr/heretic-coder-v1-GGUF
Base Architecture	Gemma 4 E4B Coder Merge (7.5B effective parameters)
Quantization	Q4_K_M (GGUF format)
File Size	5.0 GB
Active VRAM Footprint	3.27 GB (100% GPU VRAM offload)
Context Window	131,072 tokens (128K native)
Target Hardware	Consumer 8GB GPUs (AMD RX 6600/7600, NVIDIA RTX 3060/4060, Apple Silicon)
Benchmarked Speed	39.2 – 40.3 tokens / second (sustained on 8GB consumer GPU with 0 throttling)
Prompt Template	Gemma Turn Format (`<
Recommended Sampling	temp 1.0, top_p 0.95, top_k 64

---

🚀 Quick Start

1. Run Directly with Ollama (One-Liner)

You can run Heretic-Coder directly via Ollama without manually downloading GGUF files:

ollama run hf.co/Hxwvr/heretic-coder-v1-GGUF:Q4_K_M

2. Manual Ollama Modelfile Setup

If you downloaded the .gguf file locally:

FROM ./gemma-4-E4B-it-heretic-mythos-v1-Q4_K_M.gguf

PARAMETER temperature 1.0
PARAMETER top_p 0.95
PARAMETER top_k 64
PARAMETER stop "<|turn>"
PARAMETER stop "<turn|>"
PARAMETER stop "<bos>"
PARAMETER stop "<eos>"

TEMPLATE """{{- if .System }}<|turn>system
{{ .System }}<turn|>
{{- end }}
{{- range .Messages }}
{{- if eq .Role "user" }}<|turn>user
{{ .Content }}<turn|>
{{- else if eq .Role "assistant" }}<|turn>model
{{ .Content }}<turn|>
{{- end }}
{{- end }}<|turn>model
"""

Build and run:

ollama create heretic-coder:v1 -f Modelfile
ollama run heretic-coder:v1

3. llama.cpp Server

llama-server \
  -m gemma-4-E4B-it-heretic-mythos-v1-Q4_K_M.gguf \
  -ngl 99 \
  -c 131072 \
  --port 8080 \
  --temp 1.0 --top-p 0.95 --top-k 64

---

🛠️ Benchmark & Verification

Heretic-Coder was benchmarked under real-world systems coding tasks:

• Task: Concurrent async priority worker pool with exponential backoff retries and two-stage graceful shutdown.
• Tokens Evaluated: 3,460 tokens continuous.
• Speed: 39.19 tok/s on an AMD Radeon RX 6600 (Navi 23, 8GB VRAM).
• VRAM Usage: 3.27 GB / 8.00 GB (100% in VRAM, 0 layers spilled to CPU).
• Correctness: Accurately utilized (priority, time.monotonic(), func, ...) in asyncio.PriorityQueue to ensure deterministic FIFO execution order among identical priorities without function pointer collision.

---

📜 Attribution & Licensing

• Creator / Tuner: Merged, tuned, and quantized by _wvr (Hxwvr / Hexeweavr).
• Base Model: Built upon Google Gemma open weights.
• License: Governed by the Gemma Terms of Use / Open License. Free for research, personal, and commercial development under standard Gemma terms.