商汤发布开放权重Loop-9B模型,复用层加深计算

SenseTime 于周二发布了 SenseNova Loop-9B 的权重和代码。这是一款多模态模型,通过将部分层运行多次来提升计算深度。该模型参数规模约 90 亿,可接受文本和图像输入并输出文本,支持推理、编码、数据分析和工具调用。其上下文窗口可扩展至 512,000 Token。
该模型有 32 个解码器层,但会将其共享的中间层执行两次,从而在每次前向传播中完成相当于 56 层的计算量。SenseTime 表示,在部分智能体和数据分析基准测试中,它的表现可与 Qwen3.8-27B 相竞争。官方 4 bit 量化版本将权重压缩至约 4.9 GB。根据模型文档,以 256,000 Token 上下文和 8 bit KV 缓存运行该模型约需 15 GiB 内存。[SenseTime]
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(SenseTime releases open-weight Loop-9B model that reuses layers for deeper computation)

SenseTime released the weights and code for SenseNova Loop-9B on Tuesday, a multimodal model that increases computational depth by running some of its layers more than once. The roughly nine-billion-parameter model accepts text and images, produces text, and supports reasoning, coding, data analysis and tool use. Its context window extends to 512,000 tokens.
The model has 32 decoder layers, but executes its shared middle layers twice to perform 56 layers’ worth of computation in each forward pass. SenseTime says it is competitive with Qwen3.8-27B on selected agent and data-analysis benchmarks. An official four-bit quantized version reduces the weights to about 4.9 GB. Running it with a 256,000-token context and an eight-bit KV cache requires about 15 GiB of memory, according to the model documentation. [SenseTime]
正文保留英文原文,标题与摘要由机器翻译。
原文发布于 TechNode(SenseTime releases open-weight Loop-9B model that reuses layers for deeper computation),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。