Model A
Extract invoice data
Return invoice number, vendor and total as JSON.
Structured output and the quality requirement narrow the selection.
Mindverse Research / In development
One call.
The right model.
Hydra matches your task to the right model. Quality, cost and response time guide selection — with a traceable decision for every request.
Currently: local development planner. Production access is not yet activated.
Illustrative model selection
Your request
hydra
Task + priorities + eligible models
Model A
Model B ✓
Model C
One answer. One recorded choice.
Concept illustration. No request is sent and no model is executed.
- Model pool
- 25+
- Models in the current development pool
- Internal benchmark
- 85.43 / 100
- Hydra adjustment D · September 2026
- Generation cost
- $0.02062
- Per 100 planned tasks · excludes service overhead
// 01 · Selection
Different tasks. Different models.
One model per request. Your task and priorities determine which one.
Model B
Review code
Check a retry condition for a boundary error.
One model with suitable reasoning capability analyses the code and returns its finding.
Model C
Draft a customer reply
Respond empathetically to a delayed delivery without unsupported promises.
A writing-capable model works with the supplied facts and tone requirements.
Illustrative examples with placeholder models. No live requests; the historical benchmark below is separate.
// 02 · How selection works
A route with a reason.
01 / Understand
The task sets the direction.
Task type, reasoning needs and the requested output format shape which models suit a request.
02 / Select
Your priorities guide the choice.
The current planner combines application-supplied model eligibility with development measurements for the task category. Quality, generation cost and response time can be weighted.
03 / Use one model
One request. One selected model.
The illustrated route uses one model for a direct response. Other models remain alternatives. Multi-model chains are not currently available.
04 / Trace
A decision with a reason.
Task features, selection criteria and chosen models can be traced in the local decision record.
Coordinator in Germany
Routing takes place in Germany. Selected models may run with external providers.
Open-model filter
An optional portfolio filter narrows selection using reviewed model classifications and explicit self-hosting assumptions. Model weights and licence terms need to be checked before deployment.
Privacy as a requirement
Your privacy requirements shape which routes are eligible. The current planner uses the eligibility supplied by the application.
25+ models in one pool
The development pool currently covers 29 models and 261 task combinations, with task-specific rankings and adjustable priorities.
// 03 · Internal benchmark
The numbers. With their limits.
Three Hydra configurations and six historical reference models on 100 internal tasks, evaluated with Fable 5.1. Development results from September 2026.
| Configuration | Scored / 100 | Quality / 100 | USD / 100 planned tasks |
|---|---|---|---|
| GLM 5.3 FlashHistorical reference model | 99 | 87.95 | $0.11406 |
| DeepSeek V4 Pro 0813Historical reference model | 99 | 87.72 | $0.39692 |
| Hydra adjustment DHydra configuration | 98 | 85.43 | $0.02062 |
| GPT-5.6 Luna maxHistorical reference model | 99 | 84.43 | $0.18695 |
| Hydra adjustment AHydra configuration | 99 | 84.06 | ≥ $0.03989 + ? |
| MiniMax M3Historical reference model | 99 | 83.91 | $0.20855 |
| Hydra adjustment H · OSSHydra configuration | 95 | 83.83 | ≥ $0.04837 + ? |
| GPT-OSS 120B highHistorical reference model | 95 | 82.82 | ≥ $0.14708 + ? |
| GPT-5.6 Luna lowHistorical reference model | 99 | 81.77 | $0.02974 |
Quality
Mean Fable 5.1 score over scored tasks, not all 100 planned tasks. The source chart shows descriptive 95% task-bootstrap intervals (2,000 replicates). The task set was reused; there is no independent holdout evidence.
Run conditions
A and fixed controls: 9 September; D and H: 10 September 2026. Providers, timeouts, recovery and judging differed between waves.
Costs
Reported generation costs in USD for 100 planned tasks, including recorded attempts. Excludes the coordinator, orchestration, judging and service overhead. ≥ + ? marks a lower bound with an unknown remainder, not a complete price.
Comparison limits
Nine selected configurations on a direct-response workload without tools. Unequal scored coverage, historical controls and overlapping intervals do not establish universal savings, superiority or a Pareto frontier. These are internal scores, not Artificial Analysis rankings.
04 · Evaluation & integration
Your workload. Your benchmark.
The current development version plans model selection locally. Let’s discuss the quality, cost and response times your application needs — and what a suitable evaluation should demonstrate.
Node.js / Offline
node current/head.mjs inspect
node current/head.mjs verifyRun from a local Hydra installation. These commands inspect the model pool and verify the planner offline; they do not execute models. Production access is not yet activated.
// FAQ
Frequently asked questions about Hydra
What is Hydra today?
Hydra is a routing project by Mindverse, currently in development. The current planner locally selects a model using task features, eligible models and weighted priorities. Production access is not yet activated.
Does Hydra use several models per request?
Each illustrated request currently uses one selected model. Multi-model routes with planners, parallel workers or separate verifier models are not currently available.
Can I weight quality, cost and speed?
Yes. The development planner combines your quality, generation-cost and response-time preferences with task-specific development measurements. An optional portfolio filter narrows the pool to open models under the documented classifications and self-hosting assumptions.
Does all data stay in Germany?
Routing takes place in Germany. Selected models may run with external providers. Data residency and privacy requirements therefore need to be considered when making models eligible.
How can I evaluate Hydra?
The current local installation provides offline commands to inspect the model pool and verify routing decisions. These do not execute models. Talk to us about your workload and a suitable evaluation.