How it works

How Soriku measures your models and picks one for every question.

You see the idea here, not the blueprint. What matters is what comes out of it: a score per model and per category, measured on your own hardware, and a choice you can always trace back.

Photo · Vitaly Gariev / Unsplash

The capability map

Soriku doesn't trust marketing benchmarks. When you add a model, it runs a short evaluation locally, across the categories that matter in daily work (code generation, reasoning, summarisation, translation, code review, security review, and more). The output is a scorecard per model, per category, on your hardware.

Soriku tracks 13 categories and you can run the measurement again whenever you like. The score is yours, not ours.

Engine · Capability map

CategoryChosen modelScore

Code generationqwen2.5-coder:14b0.97

Code reviewqwen2.5-coder:7b0.85

Security reviewqwen2.5-coder:7b0.83

Long contextgpt-4o · own key0.88

Creative writinggemma3:4b0.94

Instruction followingqwen3.5:4b0.91

Smart routing

Every prompt gets classified in under five milliseconds. Soriku looks at the capability map and routes to the model that scores highest for that category, subject to your constraints: local-only, cost ceiling, latency budget, preferred provider. You always see who answered.

Routing is deterministic. You can pin a category to a specific model when you want full control.

Chat

Write a short story about a lighthouse keeper who finds a letter in a bottle. Around 300 words, calm tone.

Chosengemma3:4bCreative writing

Highest measured score for creative writing on this machine.

gemma3:4b0.94
qwen3:4b0.86
llama3.1:8b0.79

3 of 9 models compared

Multi-model orchestration

For high-stakes prompts, Soriku can run two or three models in parallel and verify their answers against each other. The Conductor picks the strongest response or merges them. You get one answer back, with provenance.

Workers cost more compute, so they earn their place on security reviews and architecture decisions rather than on every question.

Workers

Is this supplier contract compatible with our GDPR obligations?

Draftsqwen2.5 · local

Lists the clauses on processing and retention.

Checksllama3.1 · local

Flags that clause 7.2 allows transfer outside the EU.

Looks for gapsmistral-large · own key

Notes there is no breach-notification term.

Conductor

Two issues to fix before signing: clause 7.2 and a missing breach-notification term.

Local and remote, one pool

Soriku doesn't know, and doesn't care, whether a model runs on your laptop or in a datacenter. The capability map treats them equally. Routing picks on score, cost, and your policy. Switch off the internet and Soriku keeps working with whatever's local.

Your API keys never leave your machine. Not even in the hosted version.

Connections

Ollama7 modelsLocalOnline

LM Studio2 modelsLocalOnline

Soriku EUHosted in the EUEUOnline

Claudeclaude-sonnetOwn keyOnline

GmailSend and readMCPAsks first

Google DriveRead onlyMCPOnline

What you get access to

The Agent Persona API, the routing daemon, the benchmark runner, they all ship in the build you download. The docs cover every endpoint if you want to build on top, and you can use the whole thing without ever opening them.

Photo · Christina @ wocintechchat.com / Unsplash

Start today, free on your own computer.

Download Soriku, connect a model and ask your first question.

Download free See pricing