BAAI
English BGE v1.5 base embedding model for feature extraction, search, and RAG retrieval workflows.
BAAI/bge-base-en-v1.5 is a focused English embedding model rather than a general-purpose generative model. Its documented 109M-parameter, 438MB footprint and FlagEmbedding integration make it technically credible for dense retrieval, semantic search, and RAG feature extraction. It is materially less capable than Claude Opus 4.8, GPT-5.5, Gemini 3.5 Flash, and GLM-4.6 on broad reasoning, coding, agentic work, and multimodal I/O; those comparisons are category-level only, since this model is not designed for those tasks. Relative to Pegasus 1.5, it has narrower I/O but a more accessible local retrieval deployment profile. Developer experience is comparatively strong: official BGE documentation, a Hugging Face model card, and FlagEmbedding/FlagAutoModel loading paths provide a clear implementation route. The model card reports MTEB retrieval and embedding evaluation, supporting task fit, but the supplied evidence does not provide independently comparable retrieval figures. Its small model size can support cost-conscious self-hosted use, although actual inference cost, hardware requirements, throughput, hosted endpoints, and SLA availability are not documented in the supplied sources. Coding scores are intentionally low, not a judgment of retrieval quality. DeepSWE and LiveCodeBench list no direct result; SWE-bench and Artificial Analysis evidence was not found. A Terminal-Bench-adjacent SkillFlow study used this encoder, but reported no significant Pass@1 improvement over its no-skill baseline, which is not a direct model benchmark. Public-preference evidence is likewise absent. Adoption is supported by its maintained toolkit and model-card presence, but no usage/download measure was supplied.
