Changelog

Context Management for Toast 1

Long tool-calling runs can fill the Toast 1 context window before the model is done. Declare context_management on a Responses or Chat Completions request and the model prunes stale tool results itself as the conversation approaches the window. It works with any tool set: hosted search tools, your own function tools, or both. Every applied edit is reported on the response.

See Manage the context window for the Chat Completions shape and what the response reports.

Hosted Search Tools for Toast 1

The Chat Completions and Responses APIs now run the full Toast 1 search harness server-side. Declare a hosted tool and the API executes the search loop against your Stores, with the exact tool schemas, defaults, and truncation behavior of the open-source toast-harness.

See the Hosted Tools guide for the tools, store scope, and streaming.

Connect to Your OpenTelemetry Provider

Export every agent action as OpenTelemetry spans to Langfuse, Braintrust, or any OTLP collector. Trace each search and tool call, debug agent behavior, and find latency bottlenecks.

Set it up under Settings → Tracing. See the Trace Export guide and the announcement.

Stream Agentic Search Traces

You can now set stream: true on an Agentic Search request to receive live server-sent trace events while the search runs.

Streaming requires search_options.agentic to be enabled. Learn more in the Agentic Search guide or see the complete Search Chunks API reference. See the announcement.

Introducing Toast 1

We released Toast 1, our specialized search model for knowledge tasks, co-designed with Mixedbread Search. It provides frontier search quality, matching or outperforming Claude Opus 5 and GPT-5.6 Sol while being up to 10× cheaper and 12× faster.

Toast 1 is available through our Chat Completions API. Read more.

Faster search with a Wholembed v3 query engine

We built a query engine specifically for Wholembed v3 and extended it to cover the long-query tail we see in production. Query encoding p50 latency is now 39% lower, reducing end-to-end search p50 latency by 21%.

See the performance results.

Faster, higher-quality reranking with mxbai-rerank-v3.1-listwise

mxbai-rerank-v3.1-listwise is now available as the default reranker. It delivers gpt-5.6-sol-level ranking quality with substantially improved latency over mxbai-rerank-v3-listwise (~25–54% faster).

Learn more in our blog.

Bring Your Own Bucket

Enterprise organizations can now keep all their content in object storage they own. With bring your own bucket, Mixedbread indexes and searches your AWS S3 bucket with ephemeral compute, retaining nothing beyond memory. Your documents and every derived artifact stay in your cloud account. Authenticate with an IAM role or access keys, optionally with your own KMS key, and contact us to enable it.

Store Chunk Grep and Listing

We added two lower-level chunk retrieval APIs:

• stores.grep for exact regex matching against chunk text and generated content.
• stores.list_chunks for retrieving chunks by metadata filters with optional numeric sorting.

Instruction-following listwise reranking with mxbai-rerank-v3-listwise

mxbai-rerank-v3-listwise is now available in preview through Mixedbread Search. It evaluates a candidate list as a whole, excellent at complex tasks like recency-aware ranking, source-priority resolution, and multi-step composite instructions. On top of Wholembed v3, it improves retrieval quality by 11% and outperforms leading pointwise rerankers at instruction following.

Learn more in our release blog post.

Agentic Search Observability

We improved observability for agentic search so you can see what the agent actually did and trace the quality of its results. Inspect agent runs directly from the dashboard to understand and tune retrieval behavior.

Auto File Contextualization

Enable with_file_context to improve retrieval for long documents (topping ConTEB with 82.1 nDCG@10). During ingestion, Mixedbread generates chunk-specific explanatory context from the original file and prepends it to each chunk before embedding. See Anthropic's introduction to contextual retrieval.

Startup Accelerator Credits

Mixedbread is partnering with the Startup Accelerator to offer credits for eligible early-stage startups. Apply through the program to get Mixedbread Search credits and build AI-native search into your product from day one.

Mixedbread Search Skill

We released the Mixedbread Search skill, an agent skill that gives your coding agent the context it needs to integrate Mixedbread perfectly into your apps. Drop it into your agent and skip the back-and-forth of piecing together SDK patterns, shortening development time from exploration to working integration.

Wholembed v3

We released Wholembed v3, our new unified omnimodal multilingual late-interaction retrieval model for Mixedbread Search. It is now the default model for new stores and improves retrieval quality across languages, modalities, and real-world search tasks, with built-in support for audio and video.

Key Features

• Higher retrieval quality: Wholembed v3 delivers state-of-the-art retrieval performance across benchmarks and real-world industrial test cases, including strong results on LIMIT and BrowseComp-Plus.
• Audio and video support: Mixedbread Search can now index and retrieve audio and video content directly, alongside text and images, in new stores by default.
• Unified omnimodal retrieval: Wholembed v3 is designed to retrieve across text, audio, and vision in one system, improving robustness on noisy, heterogeneous real-world data.

Read more in the Wholembed v3 release post.

New Cookbooks

We published three new cookbooks to help you build with Mixedbread Search. The Image Search cookbook guides you through building visual search applications. Chat with PDFs demonstrates how to create conversational interfaces over your documents. Agent Memory shows how to use Mixedbread as persistent memory for AI agents.

Agentic Search

Agentic search is now available, designed for AI agents that need to retrieve information autonomously. The rewritten query is exposed in responses so you can see how the system interprets your searches. Read more in the agentic search documentation.

Web Search

Stores now support web search as a data source. You can retrieve web results directly via the store compatible API, enabling search across both uploaded documents and web content in a single query.

Search-only API Keys

You can now create API keys with search-only scope. These restricted keys can perform searches but cannot modify stores, upload files, or access other endpoints. This is useful for client-side applications where you want to limit what the key can do.

Free Tier

We launched a free plan that lets you explore Mixedbread Search without commitment. Free tier users get up to 1000 store files and access to core platform features. Token-based pricing is now displayed on the pricing page with an interactive calculator to estimate costs.

Spending Limits

Spending limits give you control over your API costs. Set a maximum spend and the system will automatically pause requests when you approach the limit, preventing unexpected charges. Cost tracking is visible in the dashboard with breakdowns by store and product.

MCP OAuth

The MCP integration now supports OAuth authentication, making it easier to connect Claude Desktop and other MCP-compatible tools to your Mixedbread stores securely. Users can authorize access through the standard OAuth flow rather than manually configuring API keys.

Vercel Marketplace Integration

Mixedbread is now available on the Vercel Marketplace as a Vercel Native integration in the Searching (and Agents) categories. Install it to:

• Connect projects in one click: we’ll add the required MXBAI_API_KEY and MXBAI_STORE_ID env vars to your Vercel projects automatically.

• Manage from the Vercel dashboard: monitor usage and costs with unified Vercel billing and access controls alongside your other integrations. Read more about it in our Vercel integration docs.

• Ship faster with a starter: deploy our example Next.js app to see Mixedbread Search in action and customize from there.

Mixedbread remains fully available as a standalone platform. Use it directly or via Vercel, whichever fits your workflow.

Mixedbread Search Public Beta

We're excited to announce the public beta of Mixedbread Search, the easy-to-use search API built from the ground up for the AI era. It is a fully-managed search engine that allows you to upload your data and start searching in minutes.

Key Features

• AI-native: Built for the AI era, with both humans and AI in mind
• Multi-modal: Search through text, images, tables, audio, and complex layouts. Video is coming soon
• Multi-lingual: Support for 100+ languages
• Fully-managed: No complex configuration, no complex setup, no complex code
• Low latency: Because you need results now, not in 5700 milliseconds
• Meaningfully state-of-the-art Search: On realistic BrowseComp-Plus benchmarks, LLM assistants are able to reach significantly better response accuracy with Mixedbread Search over existing search systems

Ingestion Speed Optimization (fast track)

Today we're excited to announce that we've optimized the ingestion speed and concurrency of our Mixedbread Search. We can ingest 1000s of files concurrently with sub 1 second latency. This is a significant improvement over the previous 20 files concurrency limit.

Key Features

• High concurrency: 1000s of files can be ingested concurrently
• Minimal latency: most files get ingested one second after they are uploaded

Public Stores

Mixedbread public stores allow users to make stores publicly accessible, so that anyone with an API key can search them. This is useful for public documentation, public knowledge bases, and other use cases where you want to share your data with the world.

Search Latency Optimization

We've optimized the hot path for the search latency of our search system, achieving sub 90ms latency without reranking and sub 120ms with reranking. This includes embedding generation, first stage single-vector retrieval, second stage multi-vector retrieval, and (optionally) reranking.

Key Features

• Minimal latency: latency so low that it's virtually undetectable by users in most scenarios.

MaxSim CPU

MaxSim CPU is a CPU-optimized version of MaxSim, a state-of-the-art similarity search operator for late-interaction models such as ColBERT and ColPali. It powers our multi-vector retrieval system, achieving sub 5ms latency on AVX2 machines.

Key Features

• CPU-optimized: Optimized for CPU-based multi-vector retrieval.
• Low latency: Sub 5ms latency on AVX2 machines.
• High throughput: 10x speedup over existing CPU-based maxsim implementations.
• Open-source: Fully accessible and customizable.

Mixedbread MCP

Introducing the Mixedbread Model Context Protocol (MCP) server, a TypeScript-based integration that exposes powerful store capabilities to AI assistants like Claude Desktop. Built as an open standard, it enables secure and controlled access to external data sources.

Key Features

• Claude Desktop Integration: Seamless integration with Claude and other MCP-compatible AI assistants
• Store Operations: Direct interaction with stores through standardized MCP tools
• Semantic Search: Enable AI assistants to search and retrieve information using natural language
• File Management: Upload, manage, and search through documents in stores
• Secure Data Access: Controlled interface for AI systems to access external data sources

Installation

npm install -g @mixedbread/mcp

mxbai CLI

Introducing the Mixedbread CLI, a command-line interface for managing Mixedbread's services directly from your terminal. Built on top of the Mixedbread SDK, it provides efficient command-line access to all core platform features.

Key Features

• Store Management: Create, list, update, and manage stores with comprehensive control
• Store File Upload & Processing: Upload files with intelligent processing strategies, metadata, and batch operations
• Semantic Search: Search through stores using natural language queries with advanced filtering
• Question Answering: Get AI-powered answers based on your store content
• Intelligent Sync: Sync files with change detection and smart processing strategies to stores

Installation

npm install -g @mixedbread/cli

Platform Alpha

Today we're excited to announce the alpha of the Mixedbread platform, bringing together all our capabilities in a unified API. Experience state-of-the-art embeddings, reranking, document parsing, and stores in one integrated solution.

Now Available in Alpha

• Embeddings API: Transform text into semantic vectors with our award-winning models
• Reranking API: Boost search relevance with advanced cross-encoder models
• Document Parsing API: Extract LLM-ready content from PDFs, DOCX, PPTX, and more
• Stores API: Fully-managed multi-modal search with automatic ingestion pipelines

mxbai-rerank-v2

The second generation of our reranking models features reinforcement learning (GRPO), extended context handling, and support for 100+ languages. These models are 8x faster than comparable alternatives while achieving higher accuracy across all benchmarks.

Models

• mxbai-rerank-base-v2 (0.5B): Balanced performance for production use
• mxbai-rerank-large-v2 (1.5B): Maximum accuracy for critical applications

mxbai-embed-xsmall-v1

Our smallest and most efficient English embedding model. Perfect for edge deployments and resource-constrained environments, it delivers competitive performance in an extra small footprint.

Batched - Dynamic Batching Library

We've open-sourced Batched, our dynamic batching library that powers Mixedbread's inference infrastructure. Achieve up to 10x throughput improvements with minimal latency impact.

Baguetter - Retrieval Testing Framework

Introducing Baguetter, our open-source framework for testing and evaluating retrieval systems. Make your search better with comprehensive testing tools designed for real-world scenarios.

BMX Algorithm

We've developed BMX, a modern take on the classic BM25 algorithm. Combining lexical and semantic signals, BMX delivers superior hybrid search performance.

deepset-mxbai-embed-de-large-v1

In collaboration with deepset, we've released a German/English embedding model that sets new performance standards among open source alternatives.

mxbai-colbert-large-v1

Our ColBERT model brings late interaction capabilities. Built on our mxbai-embed-large-v1 architecture, it achieves state-of-the-art performance on 13 BEIR benchmarks.

mxbai-embed-large-v1

Our flagship English embedding model delivers state-of-the-art performance, outperforming closed source alternatives like OpenAI's text-embedding-v3.

mxbai-embed-2d-large-v1

The world's first 2D-Matryoshka embedding model. This innovative approach allows you to reduce both the number of layers and dimensions while maintaining competitive performance.

mxbai-rerank-v1

We're launching our first generation of reranking models, available in three sizes. These models add a powerful semantic layer to existing search systems, dramatically improving result relevance.

Models

• mxbai-rerank-xsmall-v1: Ultra-efficient for high-volume applications
• mxbai-rerank-base-v1: Balanced performance and speed
• mxbai-rerank-large-v1: Maximum accuracy for critical applications