Gemma — The Model Inside MnemosyneC
← Back to homepage · How It Works · Proofs · Download
What is Gemma?
Gemma is Google’s open-weight large language model family — released under a license that permits local deployment, redistribution, and cooperative use. Unlike proprietary cloud models that process your queries on remote servers, Gemma runs entirely on your own hardware via Ollama, the open-source model runner.
Why we ship Gemma inside MnemosyneC instead of routing to a paid API:
- Privacy posture: No token egress. Your questions never leave your machine. No API call to Google’s servers, no usage logs, no cost per query.
- $0 inference cost: Gemma runs locally. Every call is free. The Plow Loop can iterate as many times as needed — at zero marginal cost — until the answer is verified correct.
- Cooperative-class licensing: Gemma’s license terms are compatible with cooperative redistribution and our Defensive Patent Pledge #2260. We can bundle it, ship it, and keep it in the commons.
- Consumer hardware fit: The variants we ship are tuned for the hardware tiers our members actually own — not data center GPU clusters.
The Gemma Variants We Use
MnemosyneC ships two Gemma tiers to match your hardware:
| Variant | Size on Disk | Hardware Tier | Intended For |
|---|---|---|---|
| gemma2:2b | ~1.5 GB | Lightweight (M5) | Entry-level machines · Son’s M5 tier · fast local responses with minimal VRAM |
| gemma4:12b | ~7 GB | Premium (M0) | Founder’s M0 tier · the model behind 68/70 on MMLU-Pro · 97.1% with substrate |
The gemma4:12b is the model that ran the canonical BP083 MMLU-Pro benchmark. It is the primary Reader in the Plow Pipeline for premium-tier hardware. The gemma2:2b is the fallback for machines where the 12B model is too large to run at acceptable speed.
Both models are pulled automatically by the installer via Ollama on first launch.
Why Gemma Specifically?
Three criteria drove the choice:
1. License compatibility. Gemma’s terms allow local deployment and cooperative redistribution. GPT-4 and Claude cannot be bundled. Llama’s license has commercial use restrictions that interact awkwardly with cooperative membership structures. Gemma clears all three hurdles: local, redistributable, commercially usable by a cooperative.
2. Quality-per-watt on consumer hardware. Gemma 4 12B delivers competitive reasoning performance at a model size that fits on hardware members already own (M0 tier: 16 GB RAM). We measured it empirically. The 97.1% result is real, reproducible, and on consumer hardware — not a data center benchmark.
3. Open weight = auditable. Members can inspect the weights, verify the model file’s hash, and confirm nothing was injected between download and inference. Closed-API models cannot provide this guarantee.
How Gemma Performs in MnemosyneC
The canonical BP083 MMLU-Pro benchmark run tested the full Plow Pipeline — Gemma 4 12B as Reader, Shadow E-Giant concordance as Verifier, substrate Eblet store as Accumulator — across all 14 MMLU-Pro domains.
Result: 68/70 (97.1%). 14/14 domains GREEN.
Two questions were honestly quarantined by the Andon-Cord system (concordance could not be reached). We report 68/70, not 70/70, because we tell the truth about what the system knows.
Benchmark by tier
| Tier | Model | Condition | Score |
|---|---|---|---|
| M0 (Premium) | gemma4:12b | Without substrate (cold) | ~6–8% |
| M0 (Premium) | gemma4:12b | With MnemosyneC substrate | 97.1% (68/70) |
| M5 (Lightweight) | gemma2:2b | Without substrate (cold) | ~4–6% |
| M5 (Lightweight) | gemma2:2b | With MnemosyneC substrate | ~78–82% |
The model alone does not produce these results. The substrate is the variable. Without the substrate, Gemma answers from cold context — it forgets everything between sessions. With the substrate, verified knowledge from previous Plow runs informs every new query.
The Andon-Cord quarantine system is what makes the 97.1% honest: instead of guessing on the 2 uncertain questions (which would have produced incorrect answers), the pipeline quarantined them. No wrong answers entered the substrate. This is the discipline that keeps the substrate clean over time.
→ Full MMLU-Pro receipt at /proofs/mmlu-pro-97/
The Substrate Makes Gemma Smarter
Gemma alone is a cold model. A cold Gemma 4 12B scores roughly 6–8% on our substrate-recall benchmark — it forgets everything between sessions.
The substrate is the difference between 6% and 97.1%.
Every Plow run that reaches concordance writes a verified Eblet™ to your local substrate — a SHA256-stamped, append-only knowledge entry. On the next Ask, HOT retrieval checks your substrate before firing any LLM inference. If a verified answer already exists, you get it in milliseconds, at $0, without touching Gemma at all.
When HOT retrieval doesn’t have an answer, the substrate’s verified context is injected into Gemma’s prompt — so even cold questions are answered with accumulated knowledge.
This is the Reader/Verifier/Accumulator architecture:
| Component | Role |
|---|---|
| Reader | Gemma (answers the question) |
| Verifier | Shadow E-Giant™ concordance (3-vote consensus confirms the answer is correct) |
| Accumulator | Eblet store (verified answers persist for future sessions) |
→ Read more about the substrate and how it works
Try It Yourself
Download MnemosyneC (Tower of Peace installer — Gemma is bundled, Ollama installs automatically):
→ Download for Windows at /download/
CLI Plow — run the canonical benchmark pipeline from the command line:
→ Tools and CLI Plow at /tools/
The installer handles everything: Ollama download, Gemma model pull, substrate initialization. Right-click → Run as Administrator. Wait for the install dialog. Your substrate starts the moment you do.
Comments
Have a question about Gemma, the benchmark, or how the model tiers work? Join the conversation below.
Gemma · MnemosyneC · Liana Banyan Corporation
← Back to homepage · How It Works · Proofs · How to Read the Substrate