Not our model, and not an official release. These weights contain no new training. They are an arithmetic
combination of three published checkpoints:
We contributed only the merge recipe and the measurements below. This is not affiliated with or endorsed by NVIDIA or
Google.
Merge: gemma-4-31B-it + 2.5 × (Ising-Calibration-1.5-31B − gemma-4-31B).
- Ising 1.5 was fine-tuned from the pretrained Gemma 4, so it does not think. Adding its fine-tuning delta to -it
keeps thinking switched on.
- Weights are bf16 with stochastic rounding. The vision tower and projector are unchanged from -it.
merge_recipe.json has the exact revisions.
Our measurements. QCalEval is scored with NVIDIA's judge on the 127-item half we used to pick α. MATH-500 uses 60
problems at levels 3–5.
Table with columns: QCalEval (thinking), QCalEval (no thinking), MATH-500, thinks | QCalEval (thinking) | QCalEval (no thinking) | MATH-500 | thinks |
|---|
| gemma-4-31B-it | 68.5 | 65.0 | 85 | 100 % |
| Ising-Calibration-1.5-31B | 74.1 | 75.0 | 13 | 0 % |
| this merge | 73.3 | 74.1 | 85 | 100 % |
Caveat: in long multi-turn agent runs (chipcal 2q1c calibration, 24 validation chips) it does worse than
gemma-4-31B-it. About 19 % of turns emit an action without reasoning, and episodes stop early. Use it for single-turn
plot questions, not as an agent.
Usage: the same as gemma-4-31B-it: use its chat template and enable thinking.
License / terms: as for Ising 1.5, use is governed by the
OpenMDW License Agreement, version 1.1 (LICENSE, © NVIDIA). Additional
information: Apache License, Version 2.0 (Gemma 4; full text and NVIDIA's notice in NOTICE). All weight files are
modified from the originals; the config, tokenizer and chat template are unmodified copies from gemma-4-31B-it. For
intended use, limitations and ethical considerations, including NVIDIA's Model Card++ subcards, see the
original Ising model card.
Cite the original work:
QCalEval
(NVIDIA) and Gemma 4:
@misc{gemmateam2026gemma4, title={Gemma 4 Technical Report}, author={Gemma Team}, year={2026}, eprint={2607.02770}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2607.02770},}