How to Configure ODS to Use Google Gemma 4 Models: A Complete Guide

Set the MODEL_PROFILE environment variable to gemma4 before running the installer to configure Osmantic ODS (Open-Source AI Stack) to use Google Gemma 4 models instead of the default Qwen family.

ODS is an open-source AI stack that automatically selects optimized local LLMs based on your hardware capabilities. When you configure ODS to use Google Gemma 4 models, the installer downloads quantized GGUF variants specifically tuned for your GPU tier, from the lightweight E2B IT model for entry-level cards to the full 31B parameter model for enterprise hardware.

Understanding ODS Model Selection Architecture

ODS determines which Gemma 4 variant to install using two core mechanisms defined in ods/installers/lib/tier-map.sh: hardware tier detection and model profile resolution.

The Hardware Tier Detection System

The ods/scripts/detect-hardware.sh script analyzes your GPU and CPU to assign a numeric hardware tier (1-4) or named tier (e.g., NV_ULTRA, SH_LARGE). This value is stored in the TIER environment variable and determines the specific Gemma 4 model variant your system can efficiently run.

The Model Profile System

The model profile is a logical name that tells ODS which model family to deploy. According to the normalize_model_profile() function in tier-map.sh, ODS accepts gemma4, gemma, or gemma-4 as valid identifiers, normalizing them all to the canonical gemma4 string. The effective_model_profile() function then determines the final profile, defaulting to gemma4 for all non-cloud tiers when MODEL_PROFILE is set to auto.

Step-by-Step Configuration for Gemma 4

Export the Model Profile Environment Variable

Before invoking the installer, export MODEL_PROFILE to target the Gemma 4 family. This variable is read during the 01-preflight phase of installation.

export MODEL_PROFILE=gemma4

Valid values include gemma4, gemma, or gemma-4—all are normalized to the same internal profile by normalize_model_profile() in ods/installers/lib/tier-map.sh.

Override Hardware Tier Detection (Optional)

To force a specific Gemma 4 variant rather than accepting the auto-detected tier, set ODS_TIER:

export ODS_TIER=2

If omitted, ods/scripts/detect-hardware.sh populates TIER based on detected hardware (e.g., an NVIDIA RTX 4090 maps to tier 4, an Intel Arc A770 maps to tier ARC).

Run the Installer and Verify Selection

Execute the top-level installer entry point:

./ods/install-core.sh

During the preflight phase, the resolve_tier_config() function checks the effective profile. When it resolves to gemma4, it calls set_gemma4_tier_config() to assign the appropriate model. The installer logs the selection:

→ resolve_tier_config: MODEL_PROFILE_EFFECTIVE=gemma4
→ set_gemma4_tier_config: TIER=2 → LLM_MODEL=gemma-4-e4b-it

Post-installation, verify the selection by checking the environment variables written by the installer:

echo "$LLM_MODEL $GGUF_FILE"

# Output: gemma-4-e4b-it gemma-4-E4B-it-Q4_K_M.gguf

How Tier Mapping Assigns Gemma 4 Models

The set_gemma4_tier_config() function in tier-map.sh uses a case $TIER block to map hardware tiers to specific Google Gemma 4 quantized models hosted on Hugging Face:

• Tier 1 (Entry Level): gemma-4-e2b-it with gemma-4-E2B-it-Q4_K_M.gguf
• Tier 2 (Prosumer): gemma-4-e4b-it with gemma-4-E4B-it-Q4_K_M.gguf
• Tier 3 (Pro): gemma-4-26b-a4b-it with gemma-4-26B-A4B-it-Q4_K_M.gguf
• Tier 4 (Enterprise): gemma-4-31b-it with gemma-4-31B-it-Q4_K_M.gguf

Each tier receives a context window and download URL appropriate for the hardware capabilities detected or specified.

Validating Your Gemma 4 Configuration

You can preview the model selection without running the full installer by using the catalog selector CLI:

python ods/scripts/select-model.py \
  --catalog ods/config/model-library.json \
  --profile gemma4 \
  --tier 2 \
  --backend nvidia \
  --vram-mb $((24*1024)) \
  --ram-gb 64

This command mirrors the installer's logic and outputs a JSON payload confirming the selected Gemma 4 model, GGUF filename, and download URL. The test suite in ods/tests/test-tier-map.sh (lines 194-197) validates this behavior by asserting that LLM_MODEL equals gemma-4-e4b-it when MODEL_PROFILE=gemma4 and TIER=2.

Platform-Specific Configuration Notes

Linux and macOS

On Unix-like systems, set environment variables using export before executing install-core.sh. The test file ods/tests/test-tier-map.sh validates the Bash implementation of set_gemma4_tier_config(), confirming correct variable assignment for tiers 1 through 4.

Windows PowerShell

For Windows deployments, use the $env: syntax:

$env:MODEL_PROFILE = "gemma4"
$env:ODS_TIER = "1"
.\ods\install-core.ps1
Write-Host "Selected model: $env:LLM_MODEL"

# Output: Selected model: gemma-4-e2b-it

The ods/tests/test-windows-model-activation.ps1 script validates this flow using ConvertTo-ModelFromTier -Tier "T1" -ModelProfile "gemma4" and asserts that the resulting LLM_MODEL matches the expected Gemma 4 variant.

Summary

• Set MODEL_PROFILE=gemma4 (or gemma, gemma-4) before installation to switch from the default Qwen family to Google Gemma 4 models.
• Hardware tiers 1-4 map to specific Gemma 4 variants (E2B, E4B, 26B A4B, and 31B) via set_gemma4_tier_config() in tier-map.sh.
• Override auto-detection by setting ODS_TIER if you need to force a specific model variant.
• Validate selections using ods/scripts/select-model.py or by inspecting the LLM_MODEL and GGUF_FILE environment variables post-installation.
• Cross-platform support is verified by test suites for both Bash (test-tier-map.sh) and PowerShell (test-windows-model-activation.ps1) environments.

Frequently Asked Questions

What happens if I leave MODEL_PROFILE unset or set it to "auto"?

When MODEL_PROFILE is unset or set to auto, the effective_model_profile() function in ods/installers/lib/tier-map.sh defaults to gemma4 for any non-cloud tier (lines 30-34). This means ODS will automatically configure Gemma 4 models for local hardware installations without requiring explicit configuration, while cloud tiers may use different defaults.

Can I force a specific Gemma 4 model variant regardless of my hardware?

Yes. While ODS auto-detects your hardware tier in ods/scripts/detect-hardware.sh, you can override this by exporting ODS_TIER with your desired tier number (1-4) before running install-core.sh. This forces set_gemma4_tier_config() to select the corresponding model variant, though running a tier 4 model on tier 1 hardware will result in poor performance or failure.

How do I verify the integrity of the downloaded Gemma 4 GGUF files?

ODS includes ods/tests/test-gemma4-artifact-pins.py, which validates the SHA-256 hashes and download URLs for all Gemma 4 model variants. You can run this test to confirm that your downloaded gemma-4-*.gguf files match the official Hugging Face repository checksums and have not been corrupted during transfer.

Does ODS support Gemma 4 on Apple Silicon or AMD GPUs?

The ods/scripts/select-model.py script accepts a --backend flag (e.g., nvidia, amd, apple) and --host-arch flag (e.g., arm64, x86_64) to filter the model catalog. While the tier map primarily targets NVIDIA CUDA tiers, you can use the selector CLI with --backend apple and --profile gemma4 to check for compatible Metal-optimized Gemma 4 variants in the ods/config/model-library.json catalog.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}