Goals and Progress

Phase 1: Evidence

Benchmark Harness

Benchmark Gemma models across hardware tiers, backends, and quantizations. Document what actually works, how fast, and at what quality. No opinions without data.

  • Single-command benchmark runner with hardware auto-detection
  • Containerized test environment with realistic agent toolset
  • Structured result artifacts (JSON, markdown, HTML dashboard)
  • One-command gemmaclaw benchmark submit to contribute results via PR
Status: Live. gemmaclaw benchmark works end-to-end.
Phase 2: Productization

Auto-Detection and Profiles

Build the auto-detection and profile-selection tooling. Ship a gemmaclaw doctor command and tested profiles that work out of the box.

  • Hardware detection (GPU, CPU, RAM, Apple Silicon Metal)
  • Tier classification and profile selection
  • Known-issue tracking with automatic fallbacks
  • gemmaclaw setup wizard with auto and advanced modes
  • Named multi-instance agents: create, list, chat --agent, message
Status: Live. gemmaclaw setup auto-detects and provisions. Multi-instance agent management shipped.
Phase 3: Community Loop

Open Profile Registry

Open the profile registry to contributions. Users report what works on their hardware, profiles get refined, coverage grows.

  • Community benchmark submission flow (via PR)
  • Configuration matrix aggregation on this site
  • Gap detection: highlight untested hardware combos
  • Failure archetype classification for Gemma post-training feedback
Status: In progress. Submission flow works, site aggregation building.

Non-GPU Support

CPU-only is a first-class path, not a fallback afterthought. Gemma 2 and Gemma 3 run on CPU via gemma.cpp. As CPU backends add Gemma 4 support, Gemmaclaw will incorporate those profiles automatically. The goal is that someone with a laptop and no discrete GPU gets a useful local Gemma assistant.

Volunteer Project

Gemmaclaw is composed of volunteers, including Google engineers and open source community members. It is not an official Google repository. Contributions and hardware reports are welcome. See the contributing guide.