Decision Models · Jev-compatible API
A decision API: one answer per option, with probabilities, to route a ticket, approve an action or score a draft. Hosting in the EU is available, and one API covers s1-fast, s1-pro and s1-vision.
1 JevBench estimates for the same decision tasks versus a 27B reasoning LLM; not a measurement of the hosted service. Method and context →
state: Mia owns a red bicycle.
question: Which color is the bicycle?
state: a business customer asks to update an invoice address.
question: Which support queue fits?
state: a viewer deletes a workspace.
question: Is this action allowed?
Two tiers, one API
GLOBAL · PEER-TO-PEER
Our lowest price per decision, on capacity that grows with demand.
Today: spare EU capacity first, then a network of independent compute providers for keys that opt in to worldwide processing.
Next: an open peer-to-peer network where anyone with a capable GPU can contribute compute to serve open models.
Coming soonHow peer-to-peer works →EU TIER
Built for GDPR — DPA accepted at signup, prompts and answers are not stored.
Aligned with the EU AI Act — model cards, benchmarks and intended use documented.
EU requests always stay in the EU and get priority.
EU tier details →Beyond our API
Pick a model from the JevBench open-weights leaderboard and run it the way you need.
s1-fast, s1-pro and s1-vision are live today. Quyet-1.0-Large is coming soon.
Learn more →Your own private endpoint for one model, EU or global.
Learn more →Installer that checks your hardware and downloads the model. Free for individuals and small teams.
Learn more →A ready-to-run box for your office.
Learn more →What you can do with it
Send context and a named question; use the typed answer in your workflow.
Support
Map each request to the team that can handle it.
Invoice address update → billing | technical | account
Support
Flag sensitive requests using your escalation policy.
Charge dispute → escalate?
Product / Safety
Check a proposed action. Keep deterministic access controls in your application.
Viewer deletes workspace → allowed?
Quality
Grade replies or summaries and escalate uncertain cases.
Draft + rubric → score
Sales
Sort messages by buying intent using your criteria.
Rollout request → hot | warm | cold
Ops / Docs
Classify document text before handing it off.
Document → invoice | contract | other
Agents
Route between search, clarification, response and stop.
Context + tools → next step?
LLM-cost routing
Classify difficulty and stakes before selecting a model.
Request → category · difficulty · stakes
Illustrative workflows. Validate representative cases on your own data before automating decisions.
The numbers, with context
Use a decision model for narrow tasks where it meets your accuracy needs. Keep an LLM for difficult reasoning and open-ended writing.
SAME DECISION TASKS
About 0.35 s versus 6.49 s.
BENCHMARK RESULTS
The compared decision-model implementation scores below frontier LLMs on accuracy.
JevBench cost and adjusted median-latency figures are estimates for earlier model implementations (before 11 October 2026), not measurements of the hosted Decision Models service or of the current lineup. JevBench is run by the Decision Models founder. Use a frontier LLM for open-ended writing or difficult reasoning. Method and sources → · Source data
Press context
“…LLMs as we know them aren't the right solution for a lot of software because they are comparatively slow and expensive.”
Based on Perplexity Decider v1.1 · 27B · FP8 · hosted in the EU
Calibrated text decisions, with up to three questions per call. Our premium option for speed, hosting in the EU and one API.
Based on Blink v0.3 26B-A4B · NVFP4 · text
Short text decisions at low latency: one question with up to 16 options. Runs on a batched engine, so probabilities for identical requests can differ slightly.
Based on Blink v0.3 26B-A4B · NVFP4 · one image
One question with one image attached, up to 26 options.
CPU
Classifies category, difficulty and stakes to help your router choose an LLM.
Coming soon
Every response names the model version that answered in model_version and in the S1-Model-Version header. The retired ids s1-pro-1, s1-fast-1 and s1-vision-1 are answered by the current models. Each question, including its state, is limited to 8,192 input tokens. Full model list: /models · Licence.
Capability, latency and cost measurements for the current lineup are being refreshed. The benchmark results published on this site were measured on earlier implementations and are labelled that way; they are not scores of the models that now answer s1-pro, s1-fast and s1-vision. Current prices are on /pricing. Method and sources: /benchmarks · earlier measurement data (JSON): download.
The interface
Send context and typed questions: up to three for s1-pro, one for s1-fast or s1-vision. Receive a chosen answer with probability information where the model supports it.
01 / INTEGRATE
Use a named state and questions of type noul, choice or score per request.
02 / INSPECT
Read the answer, confidence or probability distribution, usage receipt and effective tier.
03 / CONTROL
Choose Peer-to-Peer or EU per API key or per request with the S1-Region header.
Two approaches
Same model profiles and API request shape. Implementations depend on available capacity. Choose per API key or per request with one header.
Built to scale worldwide. Peer-to-Peer starts on spare EU capacity. When it is busy, keys you opted in can draw on a network of independent compute providers in various countries, including outside the EU/EEA.
Today
EU requests have priority. Enable “Allow worldwide processing” for each key that may leave the EU. Without opt-in, Peer-to-Peer stays on EU capacity and may be queued or rejected when busy. Existing customers keep EU-only processing unless they opt in. Request and response content is not stored on any node.
Highest data protectionFor teams with the highest data-protection demands or strict EU regulatory requirements. EU requests run only on contracted operators inside the EU, and inference data never leaves the EU. EU requests have priority, even if a key allows worldwide Peer-to-Peer processing.
Sign-in (Google) and payments (Stripe) are handled by the sub-processors listed on our sub-processor page; they never receive your prompts. Worldwide Peer-to-Peer capacity is currently rented through Lium (lium.io), whose independent GPU providers are listed there too. The open network described above is planned and carries no customer traffic today.
Built for teams
START HERE
Quickstarts, request shapes, errors and a migration guide from Jev.
Browse docs →DATA HANDLING
Where inference runs, how data is handled and which contracts you get.
Read trust overview →LLMS + MCP
Machine-readable guides and a hosted MCP server.
Agent access →FAQ
No. Decision Models is a separate service hosting open-weight decision models behind a Jev-compatible API. Jev is a trademark of TypeSafe AI; references are nominative only. Decision Models is not affiliated with TypeSafe.
Sign in with Google, create an API key in the dashboard and send your first request. Free allowance by invitation. Top up from EUR 1. Private individuals and businesses worldwide can sign up and top up. Taxes are shown in checkout before you pay.
No. Prompts and outputs are never stored or used for training. For billing we keep usage records such as token counts.
Peer-to-Peer starts on spare EU capacity, with EU requests taking priority. When busy, Peer-to-Peer can draw on a network of independent compute providers in various countries, including outside the EU/EEA. This happens only for keys with “Allow worldwide processing” enabled. Without opt-in, Peer-to-Peer stays on EU capacity and may be queued or rejected when busy. Existing customers keep EU-only processing unless they opt in. Request and response content is not stored on any node. EU-tier requests always stay in the EU. Processing locations and transfer conditions →
Only Peer-to-Peer requests from keys you opted in may use independent compute providers outside the EU/EEA when spare EU capacity is busy. EU-tier requests always stay in the EU. The switch does not establish a legal basis for transferring personal data: do not send personal data on opted-in keys unless an adequate transfer basis applies to your use. Read the DPA and sub-processor list.
Yes, with s1-vision: one image per request (PNG, JPEG or WebP). Text-only requests are billed at the s1-vision text rate; a request that contains an image is billed at the s1-vision image rate for all its input tokens (text and image positions), because image processing takes about 3× the GPU time per token. An image counts as the input tokens the model actually reads.
No. The benchmark comparison shows lower accuracy for the decision-model profile than for frontier LLMs. Validate your own examples; use a frontier LLM or human review for open-ended writing and difficult tasks. Method and results →