groxaxo/magiseek-v1-pro

MagiSeek-Pro-V1 — OpenCode serving template

This repository contains the corrected llama.cpp chat template for the Magidonia/MagiSeek p5 Q6_K checkpoint. It is designed for OpenAI-compatible clients, especially OpenCode.

The template is important: the model emits Hermes-style <tool_call> JSON. Without this template, llama.cpp may return those calls as ordinary text instead of a structured OpenAI tool_calls response.

OpenCode / llama.cpp serving

llama-server \
  --model p5.Q6_K.gguf \
  --host 0.0.0.0 --port 12434 \
  -ngl all -c 60000 -fa on \
  --jinja --chat-template-file chat_template.jinja \
  --reasoning-format deepseek \
  -ctk q8_0 -ctv q8_0 \
  --repeat-penalty 1.15 --repeat-last-n 256 \
  --presence-penalty 0.1 -np 1

For agentic use, use a sampling temperature around 0.6–0.7. Avoid temperature: 0 for long tool-use sessions because deterministic decoding can repeat read-only commands.

OpenCode connection

Configure an OpenAI-compatible provider with:

Base URL: http://HOST:12434/v1
Model:    magidonia-p5-q6-coding

The live proxy may expose these variants:

• magidonia-p5-q6-instruct — normal responses
• magidonia-p5-q6-coding — tool use and reasoning
• magidonia-p5-q6-thinking — reasoning-focused responses

Tool calls should arrive as structured OpenAI tool_calls; reasoning should arrive in reasoning_content, not leak into ordinary response content.

Files

• chat_template.jinja — corrected OpenCode-compatible llama.cpp template.

This repository publishes the serving template and instructions; model weights are not duplicated here by this template-only upload.