Best Embedding Models in 2025: What Developers Actually Use

Embedding models sit at the foundation of modern AI applications, powering everything from semantic search to retrieval-augmented generation (RAG). A recent Hacker News discussion asking developers to name their preferred embedding models has surfaced a rich landscape of options — and some surprising consensus about what actually works in production.

The conversation reveals that choosing the right embedding model is far from straightforward. Cost, latency, dimensionality, and domain specificity all play critical roles in the decision, and what works for one use case may fail spectacularly in another.

Key Takeaways From the Developer Community

• OpenAI's text-embedding-3-large remains a popular default, but developers increasingly question whether it justifies the API cost
• Open-source models from Hugging Face — particularly those on the MTEB leaderboard — are closing the gap fast
• Cohere's Embed v3 earns praise for multilingual tasks and enterprise-grade reliability
• BGE (BAAI General Embedding) models are a favorite among self-hosting enthusiasts
• Sentence-transformers library continues to be the go-to framework for running local models
• Nomic Embed and Jina Embeddings v2 are gaining traction as strong mid-tier alternatives

OpenAI Still Leads but Faces Growing Competition

OpenAI's embedding API has long been the default starting point for developers building semantic search and RAG pipelines. The company's text-embedding-3-small model costs just $0.02 per million tokens, while the larger variant — text-embedding-3-large — runs $0.13 per million tokens. Both support flexible dimensionality, letting developers trade accuracy for storage efficiency.

However, many developers in the discussion note that OpenAI's dominance is no longer a given. The proprietary nature of the API means every query incurs cost and latency, which compounds quickly at scale. For startups processing millions of documents, even fractions of a cent per query add up to significant monthly bills.

Compared to earlier models like text-embedding-ada-002, the v3 family delivers measurably better performance on benchmarks like MTEB. But the improvement gap between OpenAI and top open-source alternatives has narrowed considerably throughout 2024 and into 2025.

Open-Source Models Dominate the Self-Hosting Conversation

The most passionate recommendations in the thread center on open-source embedding models that developers can run on their own infrastructure. This eliminates per-query API costs and keeps sensitive data in-house — 2 concerns that resonate strongly with enterprise developers.

BGE-large-en-v1.5 from the Beijing Academy of Artificial Intelligence (BAAI) consistently ranks among the top models on the Massive Text Embedding Benchmark (MTEB). It produces 1,024-dimensional vectors and runs comfortably on a single GPU with 8GB of VRAM. Developers report excellent performance for English-language retrieval tasks.

For those seeking even better quality, BGE-M3 supports over 100 languages and handles documents up to 8,192 tokens — a significant upgrade over older models capped at 512 tokens. This extended context window makes it particularly useful for embedding longer documents without chunking.

Other frequently mentioned open-source options include:

• E5-large-v2 from Microsoft, known for strong zero-shot performance
• GTE-large from Alibaba's DAMO Academy, competitive on MTEB benchmarks
• Nomic-embed-text-v1.5, which offers a generous open-source license and 8,192-token context
• Jina-embeddings-v2-base-en, optimized for long documents with an 8K context window

The MTEB Leaderboard Shapes Developer Choices

The Massive Text Embedding Benchmark has become the de facto standard for comparing embedding models. Hosted on Hugging Face, it evaluates models across 8 task categories: classification, clustering, pair classification, reranking, retrieval, semantic textual similarity (STS), summarization, and bitext mining.

Developers in the discussion frequently reference MTEB scores when justifying their choices. However, several experienced practitioners caution against treating the leaderboard as gospel. Models that score highest on MTEB don't always perform best on domain-specific tasks.

'MTEB is a great starting point, but you need to evaluate on your own data,' is a sentiment echoed repeatedly. The benchmark skews toward academic datasets, which may not reflect the messy, noisy text that production systems encounter daily. Developers working in specialized domains — legal, medical, financial — consistently report that fine-tuned smaller models outperform generic large ones.

This insight highlights a broader truth: embedding quality is task-dependent. A model that excels at semantic similarity may underperform at retrieval, and vice versa.

Cohere and Voyage AI Carve Out Enterprise Niches

Beyond OpenAI and open-source options, several API-based providers earn strong recommendations. Cohere's Embed v3 stands out for its multilingual capabilities and its support for different input types — search documents, search queries, classification, and clustering each get optimized handling.

Cohere prices its embedding API competitively, and the company's focus on enterprise features like data privacy agreements and SOC 2 compliance makes it attractive for regulated industries. Developers report that Embed v3 performs particularly well for multilingual retrieval tasks, outperforming OpenAI's offerings in non-English languages.

Voyage AI, a lesser-known player, also garners enthusiastic endorsements. The company offers domain-specific models — voyage-code-2 for code search and voyage-law-2 for legal documents — that outperform general-purpose embeddings in their respective domains. Pricing starts at $0.10 per million tokens, positioning it between OpenAI's small and large model tiers.

Google's Gecko embedding model, available through the Vertex AI platform, rounds out the major API options. It produces compact 768-dimensional vectors and integrates naturally with other Google Cloud services.

Practical Guidance: How to Choose the Right Model

The discussion yields a practical decision framework that developers can follow when selecting an embedding model. The choice ultimately depends on 4 key factors: budget, latency requirements, data sensitivity, and domain specificity.

For rapid prototyping, OpenAI's text-embedding-3-small offers the fastest path to a working system. The API is well-documented, latency is low, and the cost is negligible at small scale.

For production systems processing millions of documents, self-hosted open-source models like BGE-M3 or Nomic Embed deliver better economics. The upfront investment in GPU infrastructure pays for itself within weeks for high-volume applications.

For specialized domains, fine-tuning a base model on domain-specific data consistently outperforms using a larger generic model. The sentence-transformers library makes this process accessible, requiring as few as 1,000 labeled pairs to see meaningful improvements.

Key considerations developers should evaluate:

• Dimensionality: Higher dimensions (1,024-4,096) capture more nuance but increase storage and search costs
• Context window: Models supporting 8K+ tokens reduce the need for complex chunking strategies
• Quantization support: INT8 or binary quantization can reduce storage by 4-32x with minimal quality loss
• Matryoshka representation: Some newer models let you truncate dimensions at inference time for flexible quality-cost tradeoffs
• Update frequency: API-based models can change without notice, while self-hosted models remain stable

The Rise of Late Interaction and Multi-Vector Models

A notable thread within the discussion highlights growing interest in ColBERT-style late interaction models. Unlike traditional single-vector embeddings, ColBERT produces a set of vectors — one per token — enabling more fine-grained matching at the cost of increased storage.

ColBERTv2 and its derivatives like JaColBERT show significant retrieval quality improvements over single-vector approaches, particularly for complex queries. The tradeoff is storage: a document that requires a single 1,024-dimensional vector in a standard model might need dozens of vectors in a ColBERT setup.

Projects like RAGatouille have simplified ColBERT deployment, making it accessible to developers without deep information retrieval expertise. Several commenters report 10-15% improvements in retrieval accuracy after switching from single-vector to late interaction models.

What This Means for AI Builders

The embedding model landscape in 2025 is more competitive and nuanced than ever. Developers no longer need to default to OpenAI — strong open-source alternatives exist for nearly every use case, and specialized API providers offer compelling options for domain-specific needs.

The most important takeaway from the community discussion is that evaluation on your own data matters more than any benchmark. Building a small test set of queries and relevant documents, then measuring retrieval quality across candidate models, takes hours but saves months of frustration.

Looking Ahead: Where Embedding Models Are Heading

Several trends suggest the embedding landscape will continue evolving rapidly. Matryoshka embeddings, which allow flexible dimensionality at inference time, are becoming standard in new model releases. Multimodal embeddings that handle text, images, and code in a single vector space are gaining adoption.

The convergence of embedding models with reranking pipelines — where a lightweight embedding model retrieves candidates and a more powerful cross-encoder reranks them — is emerging as the dominant architecture for production search systems. This two-stage approach delivers the best balance of cost, latency, and quality.

For developers building AI applications today, the message is clear: the best embedding model is the one that performs best on your specific data and fits your operational constraints. Start with benchmarks, but finish with your own evaluation.