Sieve 27B v2 is the stronger model on public tool benchmarks overall (91.6% vs 84.9%), and if phishing classification is your workload it is the one to beat. Mouse-1 is ahead on the two agent-shaped benchmarks, BFCL (97.6% vs 96%) and When2Call (85.8% vs 84.8%), and it is an API you can call without serving a 52 GB model.
Sieve is an open-weight decision model family from Sthānika AI. The 27B v2 release (October 18, 2026) replaced the original 27B on the Decision Index board. It is a Qwen3.8-27B fine-tune with a pointer head, calibrated with a fitted temperature, shipped with its own server.
At a glance
| Mouse-1 | Sieve 27B v2 | |
|---|---|---|
| Public tool benchmarks (avg of 4) | 84.9% | 91.6% |
| BFCL (function calling) | 97.6% | 96% |
| When2Call (call, ask or decline) | 85.8% | 84.8% |
| PhishNChips | 72.3% | 96.7% |
| How you run it | Hosted API | Self-host (~52 GB) |
| Weights | Closed | Open, Apache 2.0 |
| 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
| Sieve 27B v2 | 91.6% |
|---|---|
| Mouse-1 | 84.9% |
Function calling, exact match (%) · Higher is better
| Mouse-1 | 97.6% |
|---|---|
| Sieve 27B v2 | 96.0% |
Next API out of 53 (%) · Higher is better
| Sieve 27B v2 | 89.0% |
|---|---|
| Mouse-1 | 83.7% |
Answer, call, ask or decline (%) · Higher is better
| Mouse-1 | 85.8% |
|---|---|
| Sieve 27B v2 | 84.8% |
Phishing or legitimate email (%) · Higher is better
| Sieve 27B v2 | 96.7% |
|---|---|
| Mouse-1 | 72.3% |
Sieve's lead is concentrated in one benchmark. On PhishNChips it scores 96.7%, among the highest on the board. On API-Bank it is ahead too. On BFCL, which checks exact function calls, and When2Call, which checks whether a tool should be called at all, Mouse-1 is ahead. Those two are closest to what an agent harness asks a decision model inside its loop.
Where Mouse-1 is the better pick
- Tool calls in an agent loop. Higher on BFCL and When2Call.
- No infrastructure. One HTTPS call; nothing to provision, warm or scale.
- Production controls. Fitted confidence cutoffs, tuning on your outcomes, zero data retention by default, and a published injection result: none of Mouse-1's decisions changed under planted instructions.
How we measured
- Sieve'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 Sieve through our agent decision suites.
Sources
Checked October 20, 2026.
See also: Mouse-1 vs Decider 31B · Mouse-1 vs Darwin-27B-ZTC · All comparisons
FAQ
Is Sieve 27B better than Mouse-1?
On the four public tool benchmarks of the Jev Decision Index, Sieve 27B v2 averages 91.6% against Mouse-1's 84.9%, mostly because of PhishNChips (96.7% vs 72.3%). Mouse-1 is ahead on BFCL (97.6% vs 96.0%) and When2Call (85.8% vs 84.8%).
Is there a hosted Sieve API?
We found none as of October 20, 2026. Sthānika AI publishes Sieve 27B v2 as Apache 2.0 weights (about 52 GB) with a serving script. Mouse-1 is a hosted API at $0.10 per 1M input tokens.
Who makes Sieve?
Sthānika AI, based in India. Sieve 27B is a fine-tune of Qwen3.8-27B with a pointer head; v2 was published on October 18, 2026.