社区对比七款紧凑语言模型
AI 总结
一篇社区帖子对比了七款参数量在1.92B至9B之间的紧凑语言模型,涵盖规格、上下文长度、模态、许可证和基准测试结果。比较对象包括K2-Horizon-7B、Granite 4.2-8B、Qwen3.5-9B、MiMo-V2.6-Distill-Qwen-9B、LFM2.5-8B-A1B、Gemma 4 E4B和LFM3-VL-2B。作者指出,这些分数均由实验室报告,采用不同评测框架的结果不宜直接比较。
深度见解
该帖汇总了多款紧凑模型的规格和基准声明,为比较提供了便利起点;但由于分数来自实验室且使用了不同的评测框架,解读时仍需谨慎。
Scores are lab-reported so take them with a grain of salt.
Specs
| Model | Org | Params (total/active) | Context | Modality | License |
|---|---|---|---|---|---|
| K2-Horizon-7B | IFM | 7B dense | 512K | text | Apache-2.0 |
| Granite 4.2-8B | IBM | 8B dense (thinking) | 128K (512K ext.) | text | Apache-2.0 |
| Qwen3.5-9B | Qwen | 9B hybrid (delta-net, thinking) | 262K (~1M ext.) | text + vision + video | Apache-2.0 |
| MiMo-V2.6-Distill-Qwen-9B | Xiaomi | 9B (SFT of Qwen3.5-9B) | ~262K (base) | text + vision (visual coding) | MIT |
| LFM2.5-8B-A1B | Liquid | 8.3B / 1.5B active MoE | 128K | text | LFM 1.0 |
| Gemma 4 E4B | 4.5B effective / 8B total (PLE) | 128K | text + image + audio | Apache-2.0 | |
| LFM3-VL-2B | Liquid | 1.92B (1.73B backbone, thinking) | 128K | text + image + video | LFM 1.0 |
Text benchmarks
Superscripts: a = K2 card (IFM harness), b = Liquid harness, c = MiMo card (Xiaomi harness); no mark = own card; e = from LFM3-1.7B, the text backbone; d/† = MiMo avg@3 / internal set. Bold = best in row (cross-harness caveat applies).
部分来源内容未展示,可前往原文继续阅读。
#K2-Horizon-7B#Granite 4.2-8B#Qwen3.5-9B#MiMo-V2.6-Distill-Qwen-9B#LFM2.5-8B-A1B#Gemma 4 E4B#LFM3-VL-2B#IFM#IBM#Xiaomi#Liquid#Google