Blog.

Inside Mixedbread: How We Built Multimodal Late-Interaction at Billion Scale

Inside Mixedbread: How We Built Multimodal Late-Interaction at Billion Scale

Technical deep-dive into Mixedbread Search - the first production-ready late-interaction search with native multimodality. Learn how we achieve sub-50ms latency on billion-scale document collections.

EngineeringResearchProduct
January 21, 2026
•
11 min read
Mixedbread Team
Read More
Mixedbread now live on the Vercel Marketplace

Mixedbread now live on the Vercel Marketplace

Announcing the integration of Mixedbread Search with the Vercel Marketplace. Seamlessly integrate state-of-the-art, blazingly fast, multi-modal semantic search into your Vercel projects!

Product
3 min readOctober 23, 2025
Fantastic (small) Retrievers and How to Train Them: mxbai-edge-colbert-v0

Fantastic (small) Retrievers and How to Train Them: mxbai-edge-colbert-v0

Introducing our new family of extremely efficient ColBERT models, to serve as backbones for modern late interaction research while outperforming ColBERTv2 with just 17 million parameters.

Research
11 min readOctober 16, 2025
A Delicious Free Lunch: Better Projections Improve ColBERT

A Delicious Free Lunch: Better Projections Improve ColBERT

Discussing the unique learning constraints introduced by the MaxSim operator, and demonstrating that simple architecture improvements to accommodate for these limitations can increase performance in a free-lunch fashion.

Research
10 min readOctober 15, 2025
Introducing Mixedbread Search

Introducing Mixedbread Search

Introducing Mixedbread Search, the first search API built from the ground up for the AI era, with both humans and AI in mind. Natively multi-modal and multi-lingual, Mixedbread Search eliminates all friction and lets you find information where it lives.

Product
4 min readOctober 1, 2025
Our Research Vision, Part 1

Our Research Vision, Part 1

Introducing our research vision and our ultimate goal: Solving Search. In this blog post, we lay out our view on Mixedbread as a research lab, our approach for prioritizing our research and making it impactful for our journey toward the objective, as well as our thoughts on balancing open- and closed-source research.

Research
10 min readSeptember 24, 2025
maxsim-cpu: Maximising Maxsim Efficiency

maxsim-cpu: Maximising Maxsim Efficiency

Introducing maxsim-cpu, a much faster way to compute the late interaction's MaxSim operator on modern CPU hardware, optimised for both x86 and Mac ARM.

ResearchEngineering
9 min readJuly 15, 2025
The Hidden Ceiling: How OCR Quality Limits RAG Performance

The Hidden Ceiling: How OCR Quality Limits RAG Performance

Benchmarking shows OCR errors cap text-based RAG: top OCR still misses 4-5% NDCG@5, while Mixedbread's multimodal vector store beats perfect text by 12% and recovers 70% of lost answer accuracy.

Research
22 min readMay 14, 2025
Baked-in Brilliance: Reranking Meets RL with mxbai-rerank-v2

Baked-in Brilliance: Reranking Meets RL with mxbai-rerank-v2

Introducing mxbai-rerank-v2, the second-generation reranking models from Mixedbread. They're crispier than ever—featuring reinforcement learning, multilingual support, and extended context handling for even better accuracy. They come with the same open-source flexibility under Apache 2.0, now with boosted performance for a more powerful search experience.

Research
15 min readMarch 13, 2025
Every Byte Matters: Introducing mxbai-embed-xsmall-v1

Every Byte Matters: Introducing mxbai-embed-xsmall-v1

Announcing mxbai-embed-xsmall-v1, our smallest and most efficient English embedding model optimised for retrieval tasks. It comes with competitive performance at an extra small footprint, support for long context, binary quantization and Matryoshka representation learning.

Research
7 min readOctober 14, 2024
Baking in Performance - Dynamic Batching with Batched

Baking in Performance - Dynamic Batching with Batched

Learn how dynamic batching boosts GPU efficiency in machine learning. We introduce Batched, our open-source library that makes it easy to implement dynamic batching and improve performance in your ML projects.

Engineering
7 min readSeptember 16, 2024
Getting Better with Baguetter - New Retrieval Testing Framework

Getting Better with Baguetter - New Retrieval Testing Framework

Baguetter is our new open-source retrieval experimentation framework that combines sparse, dense, and hybrid retrieval into a single, easy-to-use interface. It enables fast benchmarking, implementation, and testing of new search methods.

Engineering
7 min readAugust 23, 2024
BM𝒳: A Freshly Baked Take on BM25

BM𝒳: A Freshly Baked Take on BM25

Introducing BMX, an iteration on the industry standard BM25 search algorithm. Through the incorporation of entropy-weighted query-document similarity and weighted query augmentation, the algorithm can increase search performance on the most relevant information retrieval benchmarks.

ResearchEngineering
15 min readAugust 12, 2024
Open Source Gets DE-licious: Mixedbread x deepset German/English Embeddings

Open Source Gets DE-licious: Mixedbread x deepset German/English Embeddings

Introducing deepset-mxbai-embed-large-v1, a new open-source German/English embedding model, developed through collaboration between deepset and Mixedbread. This model sets a new performance standard among open source peers, supporting binary quantization and Matryoshka representation learning for significant cost reductions. Outperforming domain-specific alternatives in real-world applications, it offers 97%+ infrastructure cost savings through binary MRL.

Research
10 min readJuly 18, 2024
64 bytes per embedding, yee-haw 🤠

64 bytes per embedding, yee-haw 🤠

Binary MRL combines two popular approaches to deal with the scalability issues of embeddings. It helps our embedding model achieve a 64x gain in efficiency while retaining more than 90% of performance, drastically reducing infrastructure costs and enabling new applications.

Research
10 min readApril 12, 2024
ColBERTus Maximus - Introducing mxbai-colbert-large-v1

ColBERTus Maximus - Introducing mxbai-colbert-large-v1

mxbai-colbert-large-v1 is a state-of-the-art ColBERT model for reranking and retrieval tasks. It is based on the mxbai-embed-large-v1 model and achieves state-of-the-art performance on 13 publicly available BEIR benchmarks. It's available on Hugging Face.

Research
9 min readMarch 19, 2024
Open Source Strikes Bread - New Fluffy Embedding Model

Open Source Strikes Bread - New Fluffy Embedding Model

Our English embedding model provides state-of-the-art performance among other efficiently sized models. It outperforms closed source models like OpenAI's text-embedding-v3.

Research
8 min readMarch 8, 2024
Fresh 2D-Matryoshka Embedding Model

Fresh 2D-Matryoshka Embedding Model

The 2D-🪆 model introduces a novel approach that enables you to reduce both the number of layers and the dimensions of embeddings within the model. This dual reduction strategy allows for a more compact model size while still delivering competitive performance compared to leading models, such as Nomic's embedding model. Specifically, reducing the model's layers by approximately 50% retains up to 85% of its original performance, even without additional training.

Research
11 min readMarch 4, 2024
Boost Your Search With The Crispy Mixedbread Rerank Models

Boost Your Search With The Crispy Mixedbread Rerank Models

Introducing Mixedbread rerank models - Upgrade your search results with our new, open-source reranking models from Mixedbread. These models, available in three sizes, make it easier to find relevant results by adding a semantic layer to existing search systems. They're simple to use, work with your current setup, and are proven to boost performance with many traditional and semantic search models. Check them out for a more accurate, efficient search experience.

Research
14 min readFebruary 29, 2024