ollama SLM Lab for OpenCode

A specialized test bed for optimizing Small Language Models (SLMs) on Ollama for OpenCode. This project provides a build environment to harden models like Granite and Qwen against the specific constraints of the Ollama API and OpenCode's agentic requirements.

Quick Start

Prerequisites: Install GNU Make and Ollama.

• Windows: Install via Scoop:

scoop install make ollama

• Linux/macOS: Use your standard package manager (e.g., apt install make or brew install make).

Build Models

# Ensure Ollama is running
ollama serve &

# Build optimized Granite variants
cd granite3.3/ollama
make ollama.recreate

# Build optimized Qwen variants
cd ../../qwen2.5-coder/ollama
make ollama.recreate

The Problem

Standard SLMs often fail in OpenCode environments due to:

• System Message Bloat: Large system prompts exceeding standard SLM context windows.
• Tool-Call Fragility: Inconsistent formatting between Ollama templates and model expectations.
• Tag Stripping: The Ollama API failing to parse non-JSON tags (e.g., <tool_call>), breaking the agent loop.

flowchart LR
    subgraph opencode [OpenCode]
        message[System & User Messages]
    end

    subgraph ollama [Ollama Environment]
        api[Ollama API]
        subgraph Model
            Template[Custom Jinja2 Template]
            LLMs[SLM Engines]
        end
    end

    opencode <--> api
    api <--> Template
    Template <--> LLMs

Loading

This lab bridges the gap by injecting custom Modelfiles and Jinja2 templates into the Ollama environment.

Model Specifications

Feature	Granite 3.3	Qwen 2.5-Coder
Base Model	granite3.3:latest	qwen2.5-coder:latest
Temperature	0 (Deterministic)	0 (Deterministic)
Default Context	32,768	12,288
OpenCode /init	✘ Failed	✘ Failed

Build System

We automate model creation across two variants: Fast (16k context) and Agent (64k context).

Note

Automatic Modelfile Generation: If a base .Modelfile is missing, the script automatically generates it using ollama show --modelfile.

Makefile Commands

Command	Action
make ollama.create	Build all variants (Base, Fast, Agent).
make ollama.recreate	Full clean and rebuild.
make ollama.clean	Remove temporary build files.
make ollama.mostlyclean	Uninstall models from Ollama.

Adding Support for New Models

1. Create directory: mkdir -p <model-name>/ollama

2. Create Makefile: Simply include the shared config:

   -include ../../include/ollama.mk
   

3. Build models: make ollama.recreate

Pro-Tip

The Agent version uses a 64k context window. Ensure your hardware has sufficient VRAM to handle the KV cache; otherwise, use the Fast version (16k).