ENPT

JevBench update of 17 Oct: kettei-fat leads the API and image boards, basal-omni joins ImageJevBench

Last updated:

Source · Model card: kettei

publicado (OK do Fabricio 17/10 BRT).

The JevBench maintainer (@airesearch12, 17 Oct) reports that 34 models were added or newly ranked since Thursday, naming kettei-fat as #1 on the API leaderboards and on ImageJevBench, plus Jev 1.14, Sage 1.4, Celeris-1 and StartLux v1.1. He also says Bespoke cut Nimble-v4-max to $0.04 per million input tokens and that it now enters four capped leaderboards, and that ImageJevBench has four new image-capable models. The count of 34 and the “four capped leaderboards” are his claims; we did not recount them.

What the boards show on 17 Oct (read from the pages; each board is a separate source):

  • API board: kettei-fat 77.8 composite (#1, capability 86.3, p50 0.20 s), d1-ultra 76.1, Sage 1.3.0 74.0, kettei-chub 73.5, d1 73.0; Jev 1.14.0 70.2 (#9) and Jev 1.13.0 71.5 (#7); Celeris-1 62.0 (#17); Nimble-v4-max 66.2 (#14).
  • ImageJevBench v0.3.0: kettei-fat 75.39 composite (#1), Wity-1 71.21, Imajev-4B 63.71, kettei-chub 63.35, Sage 1.4 with images 47.91 (#5, capability 75.79); basal-omni-1.0-ultra (Apache-2.0) is #8 with an artifact date of 17 Oct and Celeris-1 #9.
  • Open-weights board: StartLux-Decision v1.1 27B (Intelligence 73.2, capability 82.6) and basal-omni-1.0-ultra (74.1) are among the new entries; both are on their pages.

Related: kettei, Sage, Celeris-1, basal, StartLux-Decision.