Call closed 30 Oct 2026First cohort chosenProposals still welcome
Ensembles & Routing (Winter 2026)
Overview
A 12-week grant for researchers who work on ensembles, routing, or model composition. We cover the compute and build the engineering to run your method on current frontier models, in public. Every result is published with its full recipe and cost, and the infrastructure makes reproducing it free, so methods are comparable on the same ground and advance the Pareto frontier of accuracy and cost.
The first cohort was chosen in early November 2026. We keep reviewing applications, and as funding and capacity grow we hope to offer further grants, earlier applications first.
The problem
For a given cost or latency, what’s the best accuracy anyone can reach — and which method and combination of models gets there?
It’s one of the most burning questions in AI right now, and one of the hardest to answer. There are dozens of good models. Each is strong in different places, and prices for the same quality can differ in orders of magnitude.
We believe the strongest AI systems will be compositions of many models: an ensemble, a router, a cascade. But the evidence for it is scattered across papers, run across disparate benchmarks, and never consolidated into up-to-date proof that the frontier has to be composed. Exploring ensembles is an expensive task: every query passes through several models, so testing one on today’s frontier means a real inference bill before you see a single result. And there’s no common ground, so you rebuild the eval for each paper and still can’t compare your work to anyone else’s.
What this call seeks to accomplish
The first verified map of the quality–cost frontier for ensembles and routers, kept up to date and fully in the open.
This call attempts to remove the barriers sitting in the way. We cover your work, compute and build the tools that enable each method to be run on common ground, measured on current models across quality, cost, and latency.
Using an open-source toolkit, you can publish every result with its full recipe and full cost in tokens, dollars, and time. Every run is cached and shared, so reproducing a result is free and building on someone else’s is cheap. Entry by entry, the cohort shows everyone which combinations win at which cost and each submission pushes the frontier further.
What we provide
- Compute
- We cover up to $5,000 USD per grantee, so your ensemble can draw on as many models as it needs and properly qualify what works and doesn’t.
- Infrastructure
- Our toolkit is open source and runs on your machine. Any ensemble, router or cascade compiles to a single url4 recipe that can be rerun, shared, and published.
- Engineering
- Our engineers work alongside you, extending the toolkit around your method until it runs exactly the way you intend.
- Evaluation
- Every run is scored against an established catalogue of third-party benchmarks, including the Inspect AI evals, and feeds a shared cache that anyone running the same eval can reuse for free.
- Credit
- Recipes and code are open source, shared under Apache 2.0. You keep authorship of your ideas and are encouraged to publish or collaborate with the cohort researchers on further publications.
- Stipend
- $2,000 stipend for your time.
Who is this for
Researchers with published work on ensembles, routing, multi-agent systems, model composition, or nearby areas, and people with a well-formed idea in this space and the rigor to test it.
If you have not, but feel deeply passionate about pushing the frontier through composable intelligence, you’re welcome to apply!
Not sure your work fits? Join the OpenMined Slack, then ask in #grants and we'll help you figure it out.
What you’d do
You decide on the approach: propose a new method or reproduce a published paper on today’s models. Either way the work is open, on a public leaderboard, where every entry is a reproducible recipe. Two early milestones exist to support you:
- Week 2
- First baseline: Publish a baseline composition, so you are familiar with the toolkit before your exploration starts.
- Week 6
- First iteration of your method. Publish your first method iteration, measured against a set of models and benchmarks.
- Weeks 6–12
- Iterate and explore: Dedicated, sustained effort to refine your method, with a strong commitment toward achieving state-of-the-art or Pareto frontier results by the end of the 12 weeks.
How to apply
Propose your project. We review applications as they come in and contact you when a grant is available for it. Grants go to researchers with published work in this space, or to people with a well-formed idea about what methods they would like to try or reproduce. Nominating others is encouraged and counts in their favor.
We weigh two things most: work you have accomplished, such as a prior publication or competition write-up with code, and a specific proposal that names the method and benchmark you would like to try or the paper with the method you would like to reproduce.
How you’ll hear: we confirm every application by email. First-cohort decisions by early November 2026; after that, we write only if your application moves forward. Email grants@openmined.org if something changes on your side.
Questions before you apply? Join the OpenMined Slack and ask in #grants. The application review team will answer there.
While you wait, join the OpenMined Slack and open #fusion-research, a community of researchers working on ensembles and routing. New calls are announced in #grants.
Submitting confirms you agree with our privacy policy.
Ensembles & Routing (Winter 2026) FAQ
What can I compose with the toolkit?
You build your method using ScreamingFace Python Client by combining existing building blocks. Because the blocks nest, you are not boxed into a few fixed modes, and most published patterns map straight onto them:
- Self-consistency and majority vote, best-of-N with a judge.
- Weighted vote on a confidence signal: logprobs, entropy, verbalized confidence, or agreement.
- Ensembles with parallel members fused by a synthesizer.
- Routers that send each question to the model that fits, by rule, by embedding, or through a pre-trained router you host.
- Confidence cascades that escalate cheap-to-strong and abstain if they run out of stages.
- Cost-aware routing that takes the cheaper branch when the budget runs low.
- Corrective loops that coach a failing draft and retry.
What modalities are supported?
Text and images. You can include an image in the prompt, such as a chart, a diagram, or a page of a document, and route it through the same recipe you would use for text. Audio and video are not supported.
How is a method scored and published?
A method compiles to a url4 recipe: one readable line for the whole system that you can rerun, share, or publish, and that converts back to editable source so you can import another entry, change it, and resubmit.
Grading is completed by the engine with version-pinned benchmarks, so that all runs are comparable.
Every result carries its full cost in tokens, dollars, and latency. Every published entry ships with the url4 line and feeds into the shared cache that makes reruns close to free, and submissions bind to a cache version so you can replay any result by submission or by date.
The leaderboard marks open- and closed-weight entries and plots the Pareto frontier, the best score reachable at each cost.
Do I need my own compute budget or API keys?
No. Subsidized inference covers up to $5,000 in credits. However, the toolkit can also run locally on your keys, so using your credits is possible and allowed.
What if my reimplementation doesn’t reproduce, or the method doesn’t transfer?
That’s a finding, and we will publish it with the same care as a new record. A method that fails to transfer to current models tells the field something important. Nobody is penalized for an honest negative result, as the field needs a verified map of what has been tried, and the map needs to show which attempts lead to dead ends.
What models can I use?
Any model across the connected providers: Anthropic, OpenAI, Gemini, Hugging Face, Ollama, and OpenRouter as a broker. By the time the cohort starts, bringing a self-hosted model of your own through the same interface will also be supported.
What benchmarks are supported?
The toolkit runs a longlist from the Inspect AI eval catalogue, each pinned to an exact version so a score means the same thing on every rerun.
Currently, only single-turn benchmarks are supported, with access to retrieval tooling and coding sandboxes. For the latest, visit our leaderboard.
We encourage the cohort to tackle benchmarks which are not saturated, such as FrontierScience, MMLU, LAB-Bench, ARC-AGI-2, but you are free to choose the target benchmark.
If the benchmark your method needs isn’t in the catalogue yet, tell us and we’ll look at adding it.
What’s out of scope for a 12-week grant?
The work is zero-shot, composed from models that already exist. No pretraining or fine-tuning. It is single-turn, so no multi-turn conversation loops, tool-using agent loops - those come in later waves. Your method has to compile as a url4 recipe and run end-to-end inside 12 weeks. If it needs something the stack doesn’t support yet, tell us. We would rather extend the toolkit than turn a good idea away.
I applied and haven’t heard. What now?
Every application is confirmed by email. First-cohort decisions go out by early November 2026; if you applied after the call closed on 30 Oct 2026, your application stays open and we write only if it moves forward. Questions are welcome in #grants on the OpenMined Slack, or by email to grants@openmined.org.
Hear about future calls
We may open other themes and cohorts here. Leave your details, and we’ll be in touch when a call matches your work.
In the meantime, join the OpenMined Slack and open #fusion-research, a community of researchers working on ensembles and routing. New calls are announced in #grants.
Submitting confirms you agree with our privacy policy.