Mixedbread

Mixedbread: Search infrastructure and open embedding models for AI applications

Mixedbread is a Berlin-founded retrieval company best known for its open-source mxbai model family. Its flagship embedding model, mxbai-embed-large-v1 (Apache 2.0), has passed 5.5 million downloads on Hugging Face and posts an MTEB average of 64.68 — ahead of OpenAI's text-embedding-3-large (64.58) and Cohere's embed-english-v3.0 (64.47) in the company's published comparisons. The lineup also includes the mxbai-rerank-v2 open rerankers, a proprietary listwise reranker (mxbai-rerank-v3-listwise), and compact edge ColBERT models.

On top of the models, Mixedbread runs a managed multimodal search API: upload text, PDFs, tables, images, video, or audio, and the platform handles parsing, chunking, indexing, and reranking end to end, returning results in under 200ms across 100+ languages. There is a free Starter tier ($5 one-time credits, no card required), a $20/month Scale plan with usage-based rates on top, and custom Enterprise contracts. The platform is SOC 2 Type II and ISO 27001 certified.

Key Features

• Open-source mxbai-embed-large-v1 embedding model (Apache 2.0, 1024 dimensions, Matryoshka truncation) with 5.5M+ Hugging Face downloads
• Reranker family: open mxbai-rerank-v2 (0.5B/2B) plus the proprietary mxbai-rerank-v3-listwise, and 17M/32M-parameter edge ColBERT models
• Managed multimodal search across text, PDFs, tables, images, and video in 100+ languages with sub-200ms retrieval
• Production scale: 2.5B+ documents indexed and 10k+ queries per second in production, per the company
• Python and TypeScript SDKs, CLI, dashboard, and MCP integration for agent-facing retrieval; open-source tooling including the mgrep semantic-grep CLI (4.3k GitHub stars)
• SOC 2 Type II and ISO 27001 certified, with BYOC options at the Enterprise tier

Ideal Use Case

Mixedbread suits engineering teams building RAG pipelines or semantic search who would otherwise stitch together an embedding API, a vector database, a chunking layer, and a reranker: the hosted search product bundles that stack behind one API, while the open models serve teams that prefer to self-host their retrieval layer.

How Mixedbread differentiates

Mixedbread raised a $5.5M seed round announced in February 2025, with investors including M13, Fifth Wall, Cometa, and Endeavor Catalyst. Its credibility rests on adoption and published numbers: 5.5M+ downloads of mxbai-embed-large-v1, head-to-head MTEB results against OpenAI and Cohere, first place on the BrowseComp-Plus retrieval evaluation (self-published), and named users including Vercel, Deepset, TinyFish, and AnkiHub.

FAQ

Are the models really open source? The core embedding and rerank-v2 models are Apache 2.0 on Hugging Face; the hosted search product and v3 listwise reranker are proprietary. Is there a free tier? Yes. The Starter plan is $0 with $5 in one-time credits and no card required; paid Scale plans start at $20/month plus usage. How is it priced beyond the subscription? Usage-based: indexing from $1.50 per 1M tokens, search from $0.10 per 1K queries, storage at $0.50 per 1M tokens monthly. Does it work for agents? Yes. MCP integration, SDKs, and an agentic search mode let AI agents query stores directly with fewer tool calls.

tl;dr

Mixedbread pairs widely used open-source embedding and reranking models with a managed multimodal search API — a credible substitute for Cohere or OpenAI retrieval stacks, from a free tier to $20/month plus usage.