Key Specifications
| Vendor | google |
| Version | gemma-2b |
| Release Date | 2024-02-21 |
| Context Window | 8192 tokens |
| Input Modalities | text |
| Output Modalities | text |
| License | Gemma License |
| Documentation | https://ai.google.dev/gemma/docs |
Benchmark Performance
| Benchmark |
Score |
Unit |
Evaluated At |
Notes |
Source |
| MMLU |
48 |
% |
2024-02-21 |
5-shot |
view |
| HUMANEVAL |
43.2 |
pass@1 |
2024-02-21 |
— |
view |
| GSM8K |
27.8 |
% |
2024-02-21 |
0-shot CoT |
view |
| MATH |
27.5 |
% |
2024-02-21 |
0-shot CoT |
view |
| BBH |
53 |
% |
2024-02-21 |
3-shot CoT |
view |
| GPQA |
18 |
% |
2024-02-21 |
0-shot |
view |
| IFEVAL |
46.4 |
% |
2024-02-21 |
prompt_strict |
view |
| ARC |
77.7 |
% |
2024-02-21 |
challenge |
view |
| MUSR |
30.6 |
% |
2024-02-21 |
0-shot |
view |
| WINOGRANDE |
72 |
% |
2024-02-21 |
0-shot |
view |
Pricing
| Tier |
Price |
Currency |
| Input | $0.1 / Mtok | USD |
| Output | $0.1 / Mtok | USD |
| Cache Read | $0 / Mtok | USD |
| Cache Write | $0 / Mtok | USD |
Source:
https://ai.google.dev/pricing
· as of 2024-02-21
Compliance
- Data Residency: US
- SOC2: ✓
- HIPAA: ✗
- GDPR: ✓
- ISO 27001: ✓
Gemma 2B
Model Overview
Google Gemma 2B 轻量开源模型, 8K 上下文, 2B 参数, 适合资源受限环境与边缘部署。
Core Specifications
| Vendor |
Version |
Release Date |
Context Window |
Input Modalities |
Output Modalities |
License |
| Google |
gemma-2b |
2024-02-21 |
8K |
text |
text |
Gemma License |
| Benchmark |
Score |
Unit |
Notes |
| MMLU (Massive Multitask Language Understanding) |
48.0 |
% |
5-shot |
| HumanEval |
43.2 |
pass@1 |
— |
| GSM8K (Grade School Math 8K) |
27.8 |
% |
0-shot CoT |
| MATH |
27.5 |
% |
0-shot CoT |
| BBH (BIG-Bench Hard) |
53.0 |
% |
3-shot CoT |
| GPQA |
18.0 |
% |
0-shot |
| IFEval |
46.4 |
% |
prompt_strict |
| ARC |
77.7 |
% |
challenge |
| MUSR |
30.6 |
% |
0-shot |
| WinoGrande |
72.0 |
% |
0-shot |
Pricing
| Input |
Output |
Cache Read |
Cache Write |
| — |
— |
— |
— |
per million tokens
Strengths
- Reliable general-purpose model.
Weaknesses
- MMLU 48.0, weak knowledge reasoning.
- HumanEval 43.2, coding weak.
- Proprietary, not self-hostable.
- Context window 8K is limited.
Use Cases
References