# decisio v0.8.0 on gemma-4-31B-it (frozen, FP8 on load, one prefill per state, self-hosted) vs torchcast-decision-12b (Torchcast AI, Gemma-4-12B fine-tune, option-letter logprob readout)

As of Oct 15, 2026, decisio v0.8.0 on gemma-4-31B-it (frozen, FP8 on load, one prefill per state, self-hosted) leads torchcast-decision-12b (Torchcast AI, Gemma-4-12B fine-tune, option-letter logprob readout) in 1 of 1 shared categories: Decision models.

decisio v0.8.0 on gemma-4-31B-it (frozen, FP8 on load, one prefill per state, self-hosted): price per 1M tokens is unset in the catalog.

torchcast-decision-12b (Torchcast AI, Gemma-4-12B fine-tune, option-letter logprob readout): price per 1M tokens is unset in the catalog.

decisio v0.8.0 on gemma-4-31B-it (frozen, FP8 on load, one prefill per state, self-hosted): open weights.

torchcast-decision-12b (Torchcast AI, Gemma-4-12B fine-tune, option-letter logprob readout): open weights.

- Which to pick, Decision models: As of Oct 15, 2026, decisio v0.8.0 on gemma-4-31B-it (frozen, FP8 on load, one prefill per state, self-hosted) leads torchcast-decision-12b (Torchcast AI, Gemma-4-12B fine-tune, option-letter logprob readout) in Decision models at #4 against #35.

As of Oct 15, 2026, decisio v0.8.0 on gemma-4-31B-it (frozen, FP8 on load, one prefill per state, self-hosted) is state of the art in Decision models (since Oct 8, 2026; sources JevBench price; methodology v1.0).

As of Oct 15, 2026, torchcast-decision-12b (Torchcast AI, Gemma-4-12B fine-tune, option-letter logprob readout) is ranked #35 in Decision models (since Oct 15, 2026; sources JevBench; methodology v1.0).