Mixedbread used Prime Intellect to train a state of the art search agent that fetches relevant documents, Slack messages, Drive documents, PDFs, or Images. When paired with GPT-5.6 Sol, the trained agent outperforms Claude Fable 5, which has an accuracy of ~60%, by nearly 10 percentage points on OfficeQA Pro V2 while being 3.7x cheaper.
At a Glance
- Trained Mixedbread Toast 1, a SOTA deep search agent for knowledge intensive workflows including Finance, Legal, Healthcare, Banking, Coding, and others.
- Outperforms Claude Fable 5 by 10 absolute percentage points on OfficeQA Pro V2 while being 3.7x cheaper
- Finetuned from Qwen3.6-35B-A3B
- Mixedbread found it incredibly easy to train on their custom harness with the Prime Intellect stack
- Trained on Mixedbread's proprietary data, with Toast 1 able to generalize well beyond those domains
Mixedbread's Toast 1 Model
Knowledge work often requires hunting for important context scattered across a dozen tools. Salespeople often need to find all relevant information about a customer to catch up before an upcoming meeting. This information may be scattered between Granola which keeps meeting notes, Notion which has a bespoke CRM, multiple excel sheets that keep track of invoicing/payments, usage records, Slack messages, etc. Using a frontier model to find all the relevant context is expensive, time consuming, and prone to retrieve the wrong documents while entirely missing the right ones. Further, it leaks all this proprietary and sensitive information to the model provider.
Mixedbread builds a platform which allows its users to bring their own data and make it searchable by agents. Mixedbread Toast 1 is a specialized subagent trained to retrieve relevant context for a larger agentic model in real world use cases like Finance, Legal, Healthcare, Banking, Coding and others. While Toast 1 is trained on Mixedbread's search backend, it can be used with other search backends as well.
Training
With Prime Intellect, Mixedbread trained Toast 1 on their own harness along with their own data without any hiccups. verifiers v1 decouples the harness from the taskset and training loop, so they could plug in their existing subagent harness without rewriting it to fit the Prime Intellect stack.
That custom Mixedbread platform harness includes:
- late interaction semantic search
- grep, read and filtering on ingested content
- the ability to manage its own context
With the harness in place, prime-rl handled the training itself smoothly, letting Mixedbread focus on their data and expertise instead of fighting infrastructure bugs.
Results
Mixedbread opted to start their training with Qwen3.6-35B-A3B for its high performance floor, strong architecture to optimize for, and community adoption/support. After training, Toast 1 reaches frontier level performance at retrieving relevant context on a wide range of real world knowledge tasks, but does so 12x faster and significantly cheaper.
When combined with GPT-5.6 Sol high after training then evaluated on OfficeQA Pro V2, a benchmark for real world knowledge work tasks, Toast 1 outperforms Claude Fable by 10 percentage points at 27% of the cost. Similar performance improvements can be seen on BrowseComp-Plus and LongSeal.
The following are the results of each system in its best scoring configuration when looking at retrieval alone, averaged across the three benchmarks. Note that Toast 1 significantly pushes the pareto frontier on both cost vs performance and speed vs performance.
