Gemma is Google’s open-weight LLM family for developers—lighter-weight cousins to the closed Gemini product line. Use this lecture to practice open-vs-closed product pairing from one vendor.
Learning Objectives
By the end of this lesson, students should be able to:
- Define Gemma as Google’s open-weight developer LLM line.
- Contrast Gemma (open) with Gemini (closed product).
- Describe typical sizes and instruct variants at a high level.
- List fit cases: education, prototyping, on-prem small apps.
- State license/terms awareness for Gemma weights.
- Show a simple HF inference path.
Gemma is Google’s family of open-weight decoder-only language models intended for developers and researchers, related in branding to Gemini but released for download/adaptation under Gemma terms of use.
Architecture & Lineage
| Stage / model | Publicly known shift | Curriculum note |
|---|---|---|
| Gemma 1 | First open developer weights from Google line | 2B/7B-class teaching/deploy |
| Gemma 1.1 / instruct | Improved instruction following | Chat-style local demos |
| Gemma 2 | Stronger quality/efficiency updates | Better open baseline |
| Gemma 2 / later sizes | Expanded size menu | Still not Gemini API models |
Open vs Closed Positioning
Access
- Open weights with Gemma ToS
- Not the same as Gemini API
- HF + Kaggle distribution common
Positioning
- Developer-friendly Google open models
- Smaller than flagship closed Gemini
- Great for teaching PEFT
Vs Gemini
- Self-host possible
- Lower peak multimodal product polish
- Different terms and capabilities
Typical Use Cases
| Use case | Why this family | Watch-outs |
|---|---|---|
| Course labs / demos | Manageable sizes on one GPU | Not a Gemini replacement blindly |
| Private prototypes | On-prem experimentation | Check ToS for commercial use |
| PEFT practice | LoRA on instruct Gemma | Data quality still rules |
Engineering Touchpoint
Engineering Upsides
- Open weights from Google ecosystem
- Good teaching sizes
- Clear Gemini contrast case
Engineering Trade-offs
- Terms are not “do anything”
- Smaller ceiling than Gemini flagships
- Multimodal product features differ
“Downloading Gemma gives you Gemini.” Gemma is a separate open-weight line; Gemini remains the closed multimodal product family.
Knowledge Check
- Short Answer: Is Gemma the same as Gemini? Answer: No—Gemma is open-weight; Gemini is the closed product/API family.
- True/False: Gemma weights are intended for developers to download and adapt. Answer: True (under Gemma terms).
- Multiple Choice: Gemma is offered by: (a) Google, (b) only Anthropic, (c) only Cohere, (d) only Databricks. Answer: (a).
- Short Answer: Name a typical Gemma use in class. Answer: Single-GPU instruct demos / LoRA labs.
- True/False: Gemma ToS can still restrict some uses. Answer: True.
- Multiple Choice: Versus Gemini APIs, Gemma usually enables: (a) more self-host control, (b) identical closed SKUs, (c) no tokenization, (d) only CNNs. Answer: (a).
- Short Answer: Why mention Gemma 2 in lineage? Answer: It improved the open baseline after Gemma 1.
- True/False: You should read Gemma terms before commercial shipping. Answer: True.
- Multiple Choice: A good PEFT practice model is: (a) instruct Gemma on a GPU, (b) only random init CNNs, (c) only k-means, (d) deleting attention. Answer: (a).
- Short Answer: What lecture is next? Answer: Phi.
Key Takeaways
- Gemma = Google open-weight developer LLMs.
- Gemini = closed multimodal products/APIs.
- Ideal for labs, prototypes, and PEFT practice.
- Next: Phi.
Clarify board: Two columns—Gemma vs Gemini—fill access, modality, ops.
Lab: Run a tiny Gemma instruct generate on HF.
Recap: Gemma is Google’s open-weight counterpart to Gemini products. Continue with Phi.