basal-omni-1.0-ultra (Remek Kinas)

Cost and speed evidence

Cost per 1,000 decisions

USD 0.084647

Cost basis

measured GPU seconds x $1.19/GPU-hour

Measured latency · p50 / p95

0.240 s / 0.329 s

Adjusted latency · p50 / p95

0.630 s / 0.807 s

Latency adjustment

v1.4 local latency adjustment

Public accuracy

67.2%

Sealed accuracy

67.3%

Measurement note

Official ImageJev v0.3 S1200+P300 pass 1 on 17 Oct 2026 (image run 09:31-09:38 UTC; no new draw: the same 1,500 items as every other self-hosted row), requested by Florian. Model: Remek/basal-omni-1.0-ultra at revision b8e214cb (Apache-2.0; base google/gemma-4-31B-it). The model card calls it an experimental release trained on text decisions only, which reads images zero-shot through Gemma 4's own image encoder, without media training. Served as the model card documents: the author's own basal engine v2.1.0, unmodified (release tarball hash-verified), basal-serve --mode eager --media on, bfloat16, one NVIDIA H200 141 GB on a rented Lium spot pod (Finland, USD 3.99 per hour), offline after the weights were staged and hash-verified (network cut proven twice, before the images arrived and before the first request). Request: the item question as state and the item image as media (one image per item), one choice question with the same option text as the Jev-images rows; native option probabilities from one forward pass, no text generated. 1,500 of 1,500 items answered (0 errors), no retries. Images were sent unchanged: the harness resized and re-encoded none. 116 of the 1,275 distinct images have transparency (RGBA), and the engine flattens them by dropping the alpha channel (PIL convert('RGB')) instead of compositing them on white, so image treatment differs between rows. When a question contains Polish letters (35 items: 6 public, 29 sealed) the engine selects its Polish prompt template, as designed; this was not overridden. Latency and GPU time are the wall clock of each loopback request, one request at a time, measured on that pod (core-track median 0.24 s and 95th percentile 0.33 s before the pipeline's latency adjustment); total GPU time was 379 s for the 1,500 items. The TypeSafe-20 use-case bank was not run for this row. The author did not take part in the run.

Published 2026-10-17. Scores and ranks can change in a later release. See the method notes.