What is new
New benchmarks and notable revisions, newest first. Also available as an Atom feed.
-
Round-trip translation benchmark grows with 4 new models 11 new runs
New models include Minimax M3 (larger model entering the top 5), and also gemma-4-e2b, ling-3.0-tiny and lfm-2.5-24b-a2b (for the hardware with limited memory). Also extended existing model quants.
-
Round-trip translation benchmark v26.10.0 published
The first leaderboard of the round-trip translation benchmark is out. It compares self-hosted open models and their quantisations, with and without reasoning, and shows the memory each needs at each context size and the relative cost of each run.