Curiosity · AI Model
Gemma 2 9B
Gemma 2 9B is Google DeepMind's June 2024 open-weights small model — a 9B dense transformer with distillation-based training that topped its size class on LMSYS Chatbot Arena at release. Released under the Gemma license, it became a common starting point for on-prem single-GPU deployments.
Model specs
- Vendor
- Family
- Gemma 2
- Released
- 2024-06
- Context window
- 8,192 tokens
- Modalities
- text
- Input price
- $0.08/M tok
- Output price
- $0.08/M tok
- Pricing as of
- 2026-04-20
Strengths
- Open weights under the Gemma license — commercial use permitted
- Competitive with much larger models on LMSYS Arena at release
- Benefits from distillation from larger Gemini teachers
- Runs comfortably on a single 24GB consumer GPU at FP16
Limitations
- Only 8K context — very short by 2026 standards
- Gemma license includes responsible-use restrictions
- Superseded by Gemma 3 4B / 12B / 27B for new deployments
- Weak on coding benchmarks versus same-size coder models
Use cases
- Single-GPU production chat assistants
- Fine-tuning base for domain-specific 9B deployments
- Edge servers with modest GPU budgets
- Classification, routing, and summarization pipelines
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| MMLU | ≈71% | 2024-06 |
| HumanEval | ≈40% | 2024-06 |
| GSM8K | ≈69% | 2024-06 |
Frequently asked questions
Is Gemma 2 9B still relevant in 2026?
Mostly as a fine-tune base for legacy workloads or a cheap fallback. For new work, Gemma 3 4B or 12B offer longer context, multimodality, and updated training — with the same licensing ergonomics.
What makes Gemma 2 9B interesting?
Knowledge distillation from a larger Gemini-family teacher let the 9B student reach quality closer to 20B+ size. That made it a favorite for single-GPU fine-tuning experiments through 2024 and into 2025.
Does Gemma 2 9B support vision?
No — Gemma 2 9B is text-only. Multimodal support came later with Gemma 3 starting in 2025. If you need vision, use Gemma 3 27B or Llama 3.2 Vision instead.
Sources
- Google — Gemma 2 blog — accessed 2026-04-20
- Hugging Face — google/gemma-2-9b-it — accessed 2026-04-20