Decider 31B is the stronger model on public tool benchmarks (88.1% vs 84.9%). Mouse-1 is the one you can call today without a GPU, and it is better at deciding whether to call a tool at all: 85.8% vs 82.5% on When2Call.
Decider 31B is an open-weight decision model from Mapika on Hugging Face: a full fine-tune of Gemma 4 31B, quantized to NVFP4 and served through vLLM. It answers System One–style questions in one or two forward passes. Mouse-1 answers the same kinds of questions as a hosted API.
At a glance
| Mouse-1 | Decider 31B | |
|---|---|---|
| Public tool benchmarks (avg of 4) | 84.9% | 88.1% |
| When2Call | 85.8% | 82.5% |
| How you run it | Hosted API | Self-host, Blackwell-class GPU |
| Price | $0.10 per 1M input tokens | Your GPU time |
| Weights | Closed | Open, Apache 2.0 |
| Zero data retention | Default | Yours to run |
| Calibration fitting, fine-tuning | /v1/calibrations, /v1/tunings | Do it yourself |
| Prompt-injection results | Published: 0% flipped | None published |
Public tool benchmarks
Average of BFCL, API-Bank, When2Call and PhishNChips (%) · Higher is better
| Decider 31B | 88.1% |
|---|---|
| Mouse-1 | 84.9% |
Function calling, exact match (%) · Higher is better
| Decider 31B | 97.8% |
|---|---|
| Mouse-1 | 97.6% |
Next API out of 53 (%) · Higher is better
| Decider 31B | 85.6% |
|---|---|
| Mouse-1 | 83.7% |
Answer, call, ask or decline (%) · Higher is better
| Mouse-1 | 85.8% |
|---|---|
| Decider 31B | 82.5% |
Phishing or legitimate email (%) · Higher is better
| Decider 31B | 86.3% |
|---|---|
| Mouse-1 | 72.3% |
Decider leads on API-Bank (85.6 vs 83.7) and PhishNChips (86.3 vs 72.3), and the two are within 0.2 points on BFCL. Mouse-1 leads on When2Call. That benchmark asks whether the model should answer directly, call a tool, ask a follow-up question or decline. In an agent harness, getting that wrong means either an unneeded tool call or a missing one.
What you get beyond the model
Decider gives you weights. Mouse-1 gives you an endpoint, plus the parts around it that a harness needs in production:
- Calibrated abstention: fit a confidence cutoff on 200–400 of your own labelled decisions and get an
abstainflag on every answer. - Tuning on your outcomes: store decisions, post what actually happened, and train a version of Mouse-1 on it.
- Published injection results: on our planted-instruction test, none of Mouse-1's decisions changed. Decider's card doesn't report an injection test.
How we measured
- Decider's scores are from the Jev Decision Index board as of October 20, 2026, scored by the board's maintainers.
- Mouse-1 is not on the board. Its scores come from our own run of the board's kit, on the same rows with the same exclusions.
- We have not run Decider through our agent decision suites.
Sources
Checked October 20, 2026.
See also: Mouse-1 vs Sieve 27B · Mouse-1 vs AnyJev L0 · All comparisons
FAQ
Is Decider 31B better than Mouse-1?
On the four public tool benchmarks of the Jev Decision Index, Decider 31B averages 88.0% against Mouse-1's 84.9%, ahead on API-Bank and PhishNChips. Mouse-1 is ahead on When2Call (85.8% vs 82.5%), the benchmark for deciding whether to call a tool, ask or decline.
Can I use Decider 31B through an API?
Not as of October 20, 2026. Decider 31B is published as open weights (Apache 2.0) quantized to NVFP4, which needs a Blackwell-class GPU, and no inference provider hosts it. Mouse-1 is a hosted API at $0.10 per 1M input tokens.
What is Decider 31B built on?
It is a full fine-tune of Google's Gemma 4 31B instruction model, published on Hugging Face by Mapika in October 2026.