ENPT

Clef day two: Ollama ships it, llama.cpp support is pending, and the first independent tests arrive

Last updated:

Source

A day after launch, Cloudflare has published nothing new about Clef: no blog or changelog post, the same prices and the same Hugging Face weights. The news came from everyone else.

Ollama. PR #18741 by Jeffrey Morgan merged on 1 October (19:52 BRT), and ollama.com/library/clef (18 GB) and clef-flash (11 GB) serve the native decision head on /v1/systemone. Two loose ends: the page asks for Ollama 0.35.1, but the PR merged after the 0.35.1 tag; and the Ollama tags list a 256K context, against 64K in the model readme and 65,536 tokens on Workers AI.

llama.cpp. ggml-org published Clef-GGUF and Clef-Flash-GGUF on 2 October, marked as needing PR #29831. That PR (text only, one sequence per batch) is still open and unreviewed. The base /v1/systemone server PR merged the same morning (details). Hugging Face now lists 33 quantizations of the two sizes. Running a generic GGUF in an older Ollama gives constrained JSON output, not the decision head.

Independent tests. Nicia’s admission-decision-eval ran 85 clinical-admission cases against the hosted APIs between 1 October 18:30 and 2 October 00:10 BRT:

  • Clef caught 98–100% of cases that needed admission and admitted 97% of routine ones. Clef-flash admitted only 62–66% of routine cases.
  • At one request at a time, Jev’s median was 98–167 ms, Clef-flash 429–533 ms and Clef 638–717 ms (p95 0.9–1.1 s). Tails reached 9–26 s, two calls never returned, and about one request in five got HTTP 429 early on.
  • Clef gave the same answer on repeat calls; Jev’s probabilities moved by up to 0.12. Two Clef decisions changed when the record’s sections were reordered.

A separate 42-case test on DEV by Reid Marlow had Jev at 71.4% and Clef-flash at 66.7%, agreeing on 35 cases. Both tests are small, single-domain and run by third parties, so treat them as early signals. The hosted latency they report runs against Cloudflare’s “13× faster” claim, which compared GPU time with Jev’s API round trip.

Attention. Cloudflare’s launch post drew 610 points and 213 comments on Hacker News, and community repos (routers, eval kits, fine-tuning kits, a webcam demo) passed a dozen in a day.

What changes on the site. Clef’s scores move from 6.7 / 7.7 / 26 to 7.2 / 7.5 / 37: more integrations, more adoption, and a lower latency sub-score after the independent hosted measurements.