Mouse-1 is the more balanced model: 84.9% against 81.1% on the four public tool benchmarks, with a 21.7-point lead on phishing detection. Darwin is the better pick if When2Call-style triage is all you need and you want to host the weights yourself.
Darwin-27B-ZTC comes from FINAL-Bench. It is a 27B open-weight model that reads decisions off a zero-token head, with a fitted calibration temperature, released on Hugging Face in October 2026. Mouse-1 is a hosted decision API.
At a glance
| Mouse-1 | Darwin-27B-ZTC | |
|---|---|---|
| Public tool benchmarks (avg of 4) | 84.9% | 81.1% |
| PhishNChips | 72.3% | 50.6% |
| When2Call | 85.8% | 92.2% |
| How you run it | Hosted API | Self-host (~54 GB BF16) |
| 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
| Mouse-1 | 84.9% |
|---|---|
| Darwin-27B-ZTC | 81.1% |
Function calling, exact match (%) · Higher is better
| Mouse-1 | 97.6% |
|---|---|
| Darwin-27B-ZTC | 96.7% |
Next API out of 53 (%) · Higher is better
| Darwin-27B-ZTC | 84.8% |
|---|---|
| Mouse-1 | 83.7% |
Answer, call, ask or decline (%) · Higher is better
| Darwin-27B-ZTC | 92.2% |
|---|---|
| Mouse-1 | 85.8% |
Phishing or legitimate email (%) · Higher is better
| Mouse-1 | 72.3% |
|---|---|
| Darwin-27B-ZTC | 50.6% |
The two models fail in different places. Darwin is the best When2Call model in this comparison set (92.2%), but scores 50.6% on PhishNChips, where 50% is a coin flip. So on phishing-or-legitimate emails it is close to guessing. Mouse-1 is less extreme: lower on When2Call, much higher on PhishNChips, slightly higher on BFCL.
For an agent that reads inbound text and decides what is safe, the phishing gap is the one to weigh. It is the same kind of judgment as "is this instruction in a web page trying to steer me?"
One caution from Darwin's own card: it was trained in part on the train and dev splits of public decision benchmarks. Benchmark scores can flatter a model trained that close to the test distribution. The board also notes that Darwin's answers near a tie vary slightly between runs.
How we measured
- Darwin's scores are from the Jev Decision Index board as of October 20, 2026. Mouse-1 is not on the board; its scores come from our own run of the board's kit on the same rows.
- We have not run Darwin through our agent decision suites.
Sources
Checked October 20, 2026.
See also: Mouse-1 vs Torchcast Decision 27B · Mouse-1 vs Decider 31B · All comparisons
FAQ
Is Mouse-1 better than Darwin-27B-ZTC?
On average across the four public tool benchmarks of the Jev Decision Index, yes: 84.9% against 81.1%. Mouse-1 leads on BFCL and by a wide margin on PhishNChips (72.3% vs 50.6%). Darwin leads on When2Call (92.2% vs 85.8%) and API-Bank.
Is there a hosted API for Darwin-27B-ZTC?
We found none as of October 20, 2026. Darwin is published as Apache 2.0 weights (about 54 GB in BF16) for you to serve. Mouse-1 is a hosted API.
What is Darwin-27B-ZTC trained on?
Its model card says it is a full fine-tune of FINAL-Bench's Darwin-27B-RSI with a zero-token readout head, trained partly on the train and dev splits of public decision benchmarks.