Today, we’re opening up access to the beta version of Mixedbread Search. Mixedbread Search is the first system that lets you experience search over your own documents the way it should be: fast, accurate, multilingual, and truly multimodal. Available immediately.

mgrep is now available for opencode. It makes it 4x faster and uses 3x less tokens. simply run mgrep install-opencode

I got better quality, half the tokens, and twice the speed with @mixedbreadai 's mgrep The "trick" is a multi-vector architecture Why only now? Because the arch and quantization details are hard Join @aaxsh18 for a research talk on how it works 👇 https://t.co/tDdjimst23

opus 4.5 is impressive. anthropic forgot a bar in their report. nano banana just added it for me. the leaderboard: https://t.co/sYZyTEjEMw

we just made Claude Code - use 53% fewer tokens - respond 48% faster - give 3.2x better responses just by giving it a better grep

> be me, go to SF > never did a hackathon before > find the context engineering one from @Theoryvc > claude code yolo with Mixedbread > won > have a new PC

Mixedbread makes image search feel like magic. No keyword labeling, no manual annotation. Describe what you are looking for and instantly find it.

Mixedbread 🤝 Vercel Mixedbread Search is now available natively on the @vercel Marketplace. You can now build with our near-instant multi-modal search from the comfort of your Vercel dashboard. Search doesn't need to be complicated, sometimes everything just works.

Not sure how to go about it? Everything you need to get started can be found in this neat template: https://t.co/uXn8XT4Ept Or learn more here:

Mixedbread 🤝 Vercel Mixedbread Search is now available natively on the @vercel Marketplace. You can now build with our near-instant multi-modal search from the comfort of your Vercel dashboard. Search doesn't need to be complicated, sometimes everything just works.

Models: mxbai-edge-colbert-v0-17m: https://t.co/dlkwQRMPEL mxbai-edge-colbert-v0-32m: https://t.co/5diuidHrlZ Report: https://t.co/g2MeEi7n1q Blog post:

Despite sticking to standard practices and datasets, the resulting models are strong performers. They are some of the best long-context models around, and the 17M one matches ColBERTv2 with 1/6th the parameters and 1/3rd the embed dimension: a perfect edge device retriever!

In the process, we also conducted many ablations to understand what makes a ColBERT model better, and what does not. Among other things, we find that the embedding dimension can be lowered quite aggressively while retaining most of the model's performance.

So we decided to train one, and produce a report on the current best practices on how to take a language model all the way to being an extremely capable retrieval model. The report compiles common retrieval wisdom in one place, to give a clear view of the full process.

Scaling laws are our greatest friend when running experiments: we can test ideas out to figure out which ones are worth scaling. But for ColBERT models, we didn't know which one to use! In 2025, there still were no flash-attention, long-context small ColBERT available!

One More (Small) Thing: Introducing mxbai-colbert-edge-v0 17M and 32M. They are are the result of an easily reproducible way to train ColBERT models from scratch. They're strong, too: the 17M variant would rank first on the LongEmbed leaderboard for models under 1B parameters.

This is a Fast Rising Science post: what we hope will be a long series of papers and blog posts about narrow, focused studies that help us understand how retrieval works. Paper: https://t.co/iBIKzQM3lI Blog post:

The effects are varied, but ultimately, most projection variants are beneficial, and the better ones (which, surprisingly enough, remain very simple) introduce noticeable NDCG@10 gains at almost no cost (other than a negligible increase in parameters).

We first highlight the learning properties of MaxSim and the limitations that it comes with. Then, we build on it to propose better architectures for the final projection of ColBERT models, inspired by common deep learning feedforward block designs.

Better Projection Variants Improve ColBERT We introduce a rare free lunch: a tiny architectural modification that improves ColBERT performance across the board, without any real tradeoffs. The modification? Better final projections than the currently used linear projection.

You can now add AI Search to your @astrodotbuild Starlight 🌟 documentation "With Mixedbread, search results instantly felt more relevant, contextually accurate, and tailored to the query that the user entered." - @imax153(@EffectTS_) Read our guide how you can get started 👇