Newest Alibaba text embedding model (July 2026), built on Qwen3.7, with 256-2,560 adjustable dimensions and 128K-token inputs; Alibaba reports about 20% better MTEB multilingual, Chinese-English and code retrieval than text-embedding-v4.
Compact Nemotron 3 embedding model (about 1.1B parameters, July 2026) pruned and distilled from a Ministral 3 3B embedder, giving 2048-dimension multilingual vectors from 32K-token inputs at much lower cost than the 8B; full BF16 precision.
NVFP4-quantized build of Nemotron-3-Embed-1B for Blackwell GPUs (July 2026), retaining near-BF16 retrieval quality with the same 2048-dimension multilingual vectors and 32K-token inputs at lower memory and compute.
NVIDIA's larger Nemotron 3 text embedding model (July 2026), built on Ministral 3 8B with bidirectional attention: 4096-dimension vectors that can be truncated, 32K-token inputs, and strong multilingual and cross-lingual retrieval for RAG.
Google's first multimodal embedding model, GA in April 2026 after a March preview, mapping text, images, video, audio and PDFs into one vector space. 8,192-token input and 128 to 3,072 dimensions; incompatible with gemini-embedding-001 vectors.
NVIDIA multimodal retrieval embedder (about 1.7B parameters, December 2025) built on an Eagle VLM with Llama 3.2 1B and a SigLIP 2 vision encoder; embeds text queries and document pages as images, text or both, including tables and charts.
Lower-cost tier of Tongyi's vision-centric multimodal embedding model (October 2025), producing separate text, image and video vectors of 64-768 dimensions for cross-modal search such as text-to-image or video-to-video.
Higher-capacity Tongyi Embedding-Vision model (October 2025) generating independent text, image and video vectors of 64-1,152 dimensions; Alibaba's pick for independent-vector cross-modal retrieval, beside qwen3-vl-embedding for fused vectors.
Google DeepMind's first EmbeddingGemma (September 2025): a 308M-parameter on-device text embedding model with 768-dimension vectors truncatable to 128, 100+ languages and 2K-token inputs. Succeeded by the multimodal EmbeddingGemma 2 in October 2026.
Smallest Qwen3 text embedding model (0.6B, June 2025): up to 1024-dimension vectors with user-selectable size, 32K-token inputs, optional task instructions and 100+ natural and programming languages; the low-cost option of the 0.6B/4B/8B set.
Mid-size Qwen3 embedding model (4B) with up to 2560-dimension vectors, 32K-token inputs, instruction awareness and 100+ languages including code, balancing retrieval quality and cost between the 0.6B and 8B models.
Largest Qwen3 text embedding model (8B, June 2025), first on the MTEB multilingual leaderboard at launch; up to 4096-dimension vectors, 32K-token inputs, 100+ languages including code. Image-capable Qwen3-VL-Embedding followed in January 2026.
Qwen3-based text embedding model (June 2025) with 64-2,048 selectable dimensions and 8,192-token inputs, 15-40% better than v3 on Alibaba's evaluations. Still Alibaba's default text-embedding recommendation, though qwen3.7-text-embedding is newer.
Alias for Codestral Embed 25.05, Mistral's code-specialised embedding model for retrieval over repositories; vectors are configurable up to 3,072 dimensions and inputs up to 8K tokens.
Pinned May 2025 snapshot of Codestral Embed, which Mistral built to embed source code for semantic search and RAG; callers can request fewer output dimensions to trade retrieval quality for storage.
BAAI's LLM-based English embedding model (about 7B parameters on a Mistral backbone) that takes few-shot task examples alongside the query to adapt its embeddings in context; accepts inputs up to 8,192 tokens.
Tongyi Lab's third-generation multilingual text embedding model (2024) with 512-1,024 dimensions and 8,192-token inputs. Superseded by text-embedding-v4; Alibaba keeps it for existing v3 indexes.
Instruction-tuned variant of multilingual-e5-large: queries carry a one-sentence task description, which improves retrieval, classification and clustering across about 100 languages; 1024-dimension vectors, 512-token inputs.
BAAI's multilingual embedding model (XLM-RoBERTa-based, about 568M parameters) covering 100+ languages and inputs up to 8,192 tokens; this endpoint returns its 1024-dimension dense vectors.
BGE-M3 served with all three of its retrieval outputs per input: dense vectors, sparse lexical weights and ColBERT-style multi-vectors, for hybrid search across 100+ languages. This variant truncates inputs at 512 tokens.
