Throughput (tokens / sec)
avg. 1,920tokens/s
Top-ranked multilingual embeddings for search
Run instantly. Pay only for what you use.
Qwen3 Embedding 8B is Alibaba's flagship open-weight embedding model under Apache 2.0: 8B parameters, a 32K-token input window, and vectors of up to 4096 dimensions — with Matryoshka support down to 32, so vector size matches what your store can afford.
It is instruction-aware: a one-line task instruction steers the same model toward retrieval, classification, clustering, or code search, across 100+ natural and programming languages — no fine-tuning required.
It ranked #1 on the MTEB multilingual leaderboard at release (score 70.58) and posts 81.22 on MTEB-Code — retrieval quality that carries production RAG pipelines and semantic search.
#1MTEB multilingual at release
Score 70.58 on the leaderboard spanning retrieval, classification, and clustering.
4096embedding dimensions
Matryoshka support scales vectors down to 32 dims to cut storage and search cost.
100+languages supported
One embedding space across natural and programming languages.
Qwen3 Embedding 8B is the retrieval side of a production RAG stack — search, deduplication, and classification that stay accurate across languages and code.
Cross-language matching in 100+ languages — query in German, retrieve English documents, and vice versa.
81.22 on MTEB-Code — natural-language queries over repositories, snippets, and documentation.
A one-line task instruction tunes embeddings for retrieval, classification, or clustering — no fine-tuning required.
Truncate 4096-dim vectors down to 32 with graceful quality falloff — match vectors to your store budget.
Semantic search and RAG retrieval
MTEB-topping retrieval quality feeds cleaner context to any generation model in the catalog.
Multilingual search across 100+ languages
One embedding space for global content — query in one language, match documents in another.
Code search and documentation retrieval
81.22 on MTEB-Code — natural-language queries over repositories, snippets, and docs.
Clustering, dedup, and classification
Instruction-aware vectors adapt to the task, and Matryoshka dimensions cut storage where scale demands it.
Chat, generation, or tool calling
Outputs vectors, not text. Pair it with a generation model such as Qwen 3.6 35B-A3B for the answering side of RAG.
Inputs beyond 32K tokens
Chunk longer documents before embedding — 32K is the per-request ceiling.
Image or audio embeddings
Text only. Use a multimodal embedder for cross-modal search.
Compare Entrim’s pricing, powered by an optimized inference runtime, against other providers using the same token counts per request.
avg. 1,920tokens/s
avg. 92ms
70.58MTEB Multilingual
No. 1 on the multilingual embedding leaderboard at release.
81.22MTEB-Code
Code retrieval from natural-language queries over source and docs.
32Kinput tokens per request
Long passages embed whole — less chunking, more context per vector.
Code examples
curl -X POST "https://api.entrim.ai/v1/embeddings" \ -H "Content-Type: application/json" \ -H "Authorization: Bearer $ENTRIM_API_KEY" \ -d '{ "model": "Qwen/Qwen3-Embedding-8B", "input": "EU datacenters keep inference close to your users." }'SDKs & docs
Use familiar SDK patterns with quickstart examples for common production setups.
OpenAI-compatible
Keep the request format your team already knows.
No cold starts
Production requests are served without model spin-up delays.
Tool-heavy coding at 3B-active cost
High-throughput agentic work with 1M context
Gemma 4 quality at 4B-active speed