Zeta-Alpha-E5-Mistral
zeta-alpha-ai/Zeta-Alpha-E5-Mistral
published Aug 2024 · updated Jan 2025
Zeta-Alpha-E5-Mistral is a 7B parameter retrieval-specialized embedding model fine-tuned on E5-mistral-7b-instruct for high-quality text embeddings.
specs
| Task | Embedding / Retrieval |
| Architecture | Mistral-based (E5-mistral-7b-instruct) |
| Parameters | 7B |
| Context Length | 4096 tokens |
about this model
Zeta-Alpha-E5-Mistral is a retrieval-specialized embedding model with 7 billion parameters, fine-tuned on E5-mistral-7b-instruct. It is the first public model from Zeta Alpha's open science embedding initiative and is designed for dense passage retrieval tasks.
The model follows the instruction-tuning strategy of the original E5-mistral: queries are formatted as Instruct: <task description>\nQuery: <query>. This enables task‑specific retrieval for applications such as fact‑checking, scientific literature search, and claim verification. The model supports a maximum sequence length of 4096 tokens.
On the MTEB BE leaderboard, Zeta-Alpha-E5-Mistral achieved a top‑10 submission, demonstrating competitive retrieval performance. Additionally, Zeta Alpha provides NanoBEIR, a set of small‑scale evaluation sets derived from BEIR datasets, to facilitate rapid model evaluation.
For a complete breakdown of the training dataset and methodology, see the Zeta Alpha blog post.
best for
- ·Retrieval-augmented generation (RAG) document retrieval
- ·Semantic search in enterprise knowledge bases
- ·Claim verification and evidence retrieval
FAQ
Queries must be formatted as "Instruct: <task description>\nQuery: <query>" with an instruction prefix before the query.
Use the gigarouter OpenAI-compatible endpoint with your API key, sending a request with the model name and your input texts.
The model supports a maximum sequence length of 4096 tokens (combined query/passage and instruction).
It is optimized for retrieval tasks; for classification, consider fine-tuning or using a dedicated classifier.
We're benchmarking and onboarding Zeta-Alpha-E5-Mistral as a hosted, OpenAI-compatible API. Sign in for free credit and be ready when it lands, or tell us you want it and we'll prioritize it.