Jev alternatives that keep the SystemOne decision contract
If your stack already posts a state and typed questions to a Jev endpoint, you do not need to rewrite your schema to leave closed weights or slow p95s behind. Paste your payload below to verify compatibility with POST /api/v1/decisions.
Validated locally in your browser — nothing pasted here is sent to the server.
- Alias "jev-latest" is accepted as-is and resolves to cloudflare/clef-flash; the snippet below pins "cloudflare/clef-flash" explicitly so model selection stays auditable.
- Id-keyed questions map detected: 100% drop-in compatible with POST /api/v1/decisions.
curl --fail-with-body https://clef-flash.com/api/v1/decisions \
-H "authorization: Bearer $CLEF_API_KEY" \
-H "content-type: application/json" \
--data-binary @- <<'CLEF_DECISION'
{
"model": "cloudflare/clef-flash",
"state": {
"ticket_id": "inc_4821",
"plan": "enterprise",
"message": "Webhook deliveries to eu-central-1 started timing out after 14:05 UTC and orders are queuing."
},
"questions": {
"owner_queue": {
"type": "choice",
"instructions": "Which engineering queue owns this incident?",
"criteria": {
"platform_edge": "webhooks, networking, regional routing or TLS failures",
"billing_ledger": "invoices, charges, refunds or credit balance",
"auth_keys": "API keys, token revocation or permissions"
}
},
"page_oncall": {
"type": "noul",
"instructions": "Should on-call be paged immediately?",
"criteria": {
"true": "production traffic or paid customer orders are actively blocked",
"false": "cosmetic bug, question or non-urgent request"
}
}
}
}
CLEF_DECISION
Ship this in production with a $19 Pro key
2,000 decisions/day, 8 concurrent requests in flight, text + JSON + image states, and an API key emailed in about a minute.
Why engineering teams evaluate Jev alternatives
Jev proved that structured decisions belong in a dedicated forward pass rather than a chat completion. Once a router, fraud check or moderation gate moves into the synchronous request path, three operational questions follow: how fast the median call finishes, what a million input tokens cost at scale, and whether you can run the weights yourself if a vendor changes terms.
Cloudflare released Clef (27B) and Clef-Flash (9B) under Apache 2.0 on 1 October 2026 to answer those exact constraints. In Cloudflare’s published 43-task benchmark, Clef-Flash reached a 38.8 ms median (about 39 ms) and 122.4 ms p95 — more than ten times faster than Jev on the same decision format — while keeping open weights you can pull locally via Ollama whenever you want.
- Same SystemOne request shape: state plus choice, noul and score questions.
- Apache 2.0 open weights for Clef-Flash (9B) and Clef (27B), zero vendor lock-in.
- Upstream input billed at $0.019 per million tokens on Clef-Flash; output tokens are free.
- Explicit pre-call character ceilings so oversized states are refused instead of silently half-read.
Served SystemOne models compared side by side
Three open-weight families speak the SystemOne contract on this deployment alongside legacy Jev client aliases. Note the difference between published model cards and measured API ceilings: while the Clef-Flash model card states a 64K window, this API enforces the 24,576 tokens measured end-to-end (120,000 characters) so no tail is ever billed after silent truncation.
| Model ID | Params & licence | Served window (ceiling) | Vision | Upstream input |
|---|---|---|---|---|
| cloudflare/clef-flash | 9B · Apache 2.0 | 24,576 tokens (120,000 chars) | Yes (up to 4 images) | $0.019 / 1M |
| cloudflare/clef | 27B · Apache 2.0 | 16,384 tokens (80,000 chars) | Yes (up to 4 images) | $0.042 / 1M |
| cloudflare/clef-omni | Omni · Apache 2.0 | 65,529 tokens (200,000 chars) | Yes (up to 4 images) | $0.150 / 1M |
| jaredpalmer/kev-4b | 4B · Open weights | 8,192 tokens (20,000 chars) | No (text & JSON) | $0.042 / 1M |
What changes when migrating from Jev (and what stays identical)
A SystemOne client already sends an object state or text string plus a questions map keyed by question ID. Pointing your base URL at https://clef-flash.com/api/v1/decisions and passing Authorization: Bearer cfl_... keeps every choice, noul and score schema intact, including legacy single-label bodies carrying text, instructions and labels.
One deliberate rule protects auditability: model names like jev, jev-latest, kev and kev-latest are accepted as drop-in aliases that route to cloudflare/clef-flash, and the JSON response always names the exact model ID that answered. Sending model: "kev" will not secretly hit jaredpalmer/kev-4b — pass jaredpalmer/kev-4b by full ID when you specifically want the 4B text-only weights.
How to benchmark before switching production traffic
Never trust a generic leaderboard on your own taxonomy. Export fifty representative states from your logs, run them through the free browser playground or a $19 thirty-day Pro key (2,000 decisions per day, 8 concurrent requests in flight, no auto-renewal), and compare agreement and tail latency against your current Jev baseline.
Start on cloudflare/clef-flash for synchronous request-path work. Move individual hard questions to cloudflare/clef (27B) only where your labelled sample shows a measurable accuracy gap, or to cloudflare/clef-omni when serialized states exceed 120,000 characters.
Frequently asked questions
Do I have to rewrite my Jev questions object into an array?
No. POST /api/v1/decisions accepts both the id-keyed map ({ route: { type, instructions, criteria } }) used by Jev and Kev clients and the array form documented on this site, as well as legacy { text, instructions, labels } payloads.
What happens if my existing code sends model: "jev-latest" or "kev"?
Both aliases are accepted so an unchanged SDK can connect immediately, and both route to the default cloudflare/clef-flash deployment. The response payload always states model: "cloudflare/clef-flash" so downstream logs never confuse an alias with the weights that ran. Specify jaredpalmer/kev-4b or cloudflare/clef by full ID when you want a specific non-default model.
Why does the table list 24,576 tokens for Clef-Flash instead of 64K?
The published Clef-Flash model card states 64K tokens, whereas live endpoint measurements truncate or refuse past 24,576 tokens. Enforcing 120,000 characters (inside 24,576 tokens) before the call guarantees a 200 OK means your entire state was read.
Is the $19 Pro plan a recurring subscription?
No. Every tier on Clef-Flash is sold as a fixed 30-day or 365-day term with zero auto-renewal. You receive your cfl_ API key by email in about a minute, and if you stop renewing after a migration test nothing charges your card again.
Keep reading
- LLM router API
- AI content moderation API
- Clef vs Clef-Flash (27B vs 9B)
- API documentation
- Pricing & API keys
Model notes
One short email when a new decision model lands or a benchmark changes. No spam, unsubscribe in one click.