Why it mattersTeams building routing, policy or document-check steps for agents can test Clef beside Jev on the same local runtime, instead of writing to two separate APIs.
A team that routes a request, checks a yes-or-no condition or reads a field off a scanned form inside an agent has usually written to one vendor's API for that step. On 19 October Ollama added Clef, Cloudflare's 27B multimodal decision model, so that step can now run on a local Ollama install or on Ollama's cloud. Ollama says the model is compatible with the Jev and System One APIs, which another Ollama post has supported since 29 September, so a team already calling Jev through Ollama can point the same code at Clef and compare the two.
Clef was fine-tuned from Qwen3.8-27B and is released under the Apache 2.0 license. Ollama describes a different shape from an ordinary language model: Clef reads the input once and scores every answer to every question at the same time, instead of writing a response word by word. That shape fits the jobs Ollama lists for it: picking one destination from a set of routes, returning true or false for a policy question with a probability, reading named fields off a photo of a document, and checking a tool call before an agent runs it. Text, JSON and images all go in as input.
What the benchmarks show
Ollama reports results from Cloudflare's own announcement, scored on the Jev Decision Index. Cloudflare says Clef beats Jev on most of the suite, with the full-size Clef at 98.47 on BFCL case exact against Jev's 95.75, 94.20 on BANKING77 macro-F1 against 79.74, and 97.43 on CLINC150+OOS against 89.27. Jev holds When2Call accuracy at 80.97 against Clef's 72.37, and BRIGHT nDCG@10 at 47.52 against 45.91. On four workflow evaluations Cloudflare ran, Clef wins invoice processing, customer service and security incidents, and loses agent trace observability to Jev at 71.6 against 68.5. The numbers are the vendor's, measured against its own index.
Latency across 43 benchmarks, as the Ollama post reports: Clef median 209.3 milliseconds, p95 238.6 milliseconds. Clef-flash, the smaller variant, is 38.8 median and 122.4 at p95. Jev is 524.1 median and 536.0 at p95. Laya is the fastest at 5.8 milliseconds median, but its accuracy numbers are the lowest in every row of the table.
Ollama lists a Clef model page in its library, cloud documentation and a download link for anyone who wants to try the model locally. Pricing for Ollama's own cloud is not stated on the Clef post and sits behind the main pricing page.
The news here is a tool a team can swap in without rewriting its integration. A pipeline that already calls Jev through Ollama for routing or policy checks can call Clef through the same client and keep its own evaluation set to see which model answers its questions better. Running both models on Ollama also removes a network call out to Cloudflare's edge for teams that already keep their inference on their own machines, which is the reason to pay attention to this release even when the benchmark rows above come out even on the problem a particular team is grading.
Source
Clef is available on Ollama's cloud, Ollama Blog. The benchmark tables cite Cloudflare's own announcement.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
