A preliminary version of the smallest model in the upcoming Granite 4.0 family, released May 2025. It utilizes a novel hybrid Mamba-2/Transformer, fine-grained mixture of experts (MoE) architecture (7B total parameters, 1B active at inference). This preview version is partially trained (2.5T tokens) but demonstrates significant memory efficiency and performance potential, validated for at least 128K context length without positional encoding.
发布日期2025年5月2日
参数规模7B
上下文长度—
许可证Apache 2.0
知识截止—
Benchmarks
评测成绩
| 评测基准 | 类别 | 分数 | 来源 |
|---|---|---|---|
| AttaQ | safety | 86.1 | 来源 |
| HumanEval | reasoning code | 82.4 | 来源 |
| HumanEval+ | reasoning | 78.3 | 来源 |
| GSM8k | math reasoning | 70.1 | 来源 |
| IFEval | general | 63.0 | 来源 |
| MMLU | general reasoning language math | 60.4 | 来源 |
| TruthfulQA | general reasoning legal healthcare finance | 58.1 | 来源 |
| BIG-Bench Hard | reasoning math language | 55.7 | 来源 |
| DROP | reasoning math | 46.2 | 来源 |
| AlpacaEval 2.0 | general creativity reasoning | 35.2 | 来源 |
| Arena Hard | general reasoning creativity | 26.7 | 来源 |
| PopQA | general reasoning | 22.9 | 来源 |
Pricing
API 价格对比
暂无 API 价格。