Curiosity · AI Model
SFR-Embedding-Mistral
SFR-Embedding-Mistral, released by Salesforce Research in early 2024, is a 7B-parameter decoder-based embedding model built on Mistral 7B with the E5-Mistral instruction-tuning recipe and a curated Salesforce multi-task dataset. It topped the MTEB leaderboard at launch and helped establish the trend of LLM-based embeddings beating classical encoder models.
Model specs
- Vendor
- Salesforce
- Family
- SFR-Embedding
- Released
- 2024-01
- Context window
- 32,768 tokens
- Modalities
- text
Strengths
- Topped MTEB English at release
- Supports long 32K-token passages via Mistral base
- Open weights for research and evaluation
Limitations
- 7B size is heavy for CPU or edge deployments
- Research license — check terms before production use
- English-centric — weaker multilingual coverage
Use cases
- High-accuracy RAG pipelines on GPU infrastructure
- Semantic search over heterogeneous corpora
- Research baselines for LLM-based embedders
- Classification and clustering on long-text inputs
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| MTEB English average | ≈67 (leader at launch) | 2024-01 |
| BEIR average | leading open-weights at launch | 2024-01 |
Frequently asked questions
What is SFR-Embedding-Mistral?
SFR-Embedding-Mistral is Salesforce Research's open-weights 7-billion-parameter English embedding model, fine-tuned from Mistral 7B with an instruction-tuning recipe similar to E5-Mistral.
How does SFR-Embedding compare to NV-Embed v2?
Both are Mistral 7B-based embedders. SFR led MTEB in early 2024; NV-Embed v2 later pushed the state of the art further. Either is a strong high-accuracy choice when GPU inference is acceptable.
Can I use SFR-Embedding-Mistral commercially?
Check the model card — the release license is intended for research, and commercial use may require a separate agreement with Salesforce.
Sources
- Salesforce Research — SFR-Embedding blog — accessed 2026-04-20
- Hugging Face — Salesforce/SFR-Embedding-Mistral — accessed 2026-04-20