Published
Updated
Sthānika's Sieve-31B card prints Decision Index 0.3.2 Full score 63.8
Cuth posted Sieve 31B on October 20, 2026. The Hugging Face card prints Decision Index 0.3.2 Full score 63.8, board data of October 14, and a public index of 62.52 on the authors' own 0.3 run. Apache-2.0 LoRA on Gemma 4 31B. This desk did not run the weights.
Cuth posted Sieve 31B on October 20, 2026, at 01:50 UTC. The text names Sthānika, a lab in Hyderabad, a Full score of 63.8, second globally, and a LoRA adapter on Gemma 4 31B with weights on Hugging Face. 丰收, 01:24 UTC, prints 63.8 and second globally. The share image on this page is the still from that post.
The Hugging Face card, opened October 20, lists license apache-2.0. The adapter is LoRA rank 128 on google/gemma-4-31B-it, revision 842da37, 979M trainable parameters on all 60 text layers. Questions are choice, noul, and score. Decision Index Full score is 63.8 on edition 0.3.2, board data of 2026-10-14. The authors’ own run of the edition 0.3 public suite, official kit at 62d2f51, is public index 62.52, raw 71.08, 140,178 scoreable requests over 43 benchmarks, 37 averaged. Area skill times 100 on that public run: Knowledge and Reasoning 49.9, Language Understanding 65.6, Retrieval and Classification 69.5, Tools and Automation 78.7, Arts and Human Taste 43.5. Latency on one RTX PRO 6000, 400 random public 0.3 requests, median 163 ms. The card says the leaderboard lists 253 ms on the same GPU class. Temperature is fixed at 8.5495. Each prompt row may run to 32,768 tokens. A choice may have 255 options. Training used 2x B200 for about 12 hours, 190,314 questions. Code is sthanika-ai/Sieve. We did not load the adapter.
sthanika.ai, opened the same day, prints Sieve-31B at 63.81, tied 3rd of 174 models, against 57.62 for the same base model read the same way, and 97.8% on BFCL. Sieve-27B is 60.17, tied 5th of 149. The family is eight open-weight models from 0.8B to 31B, on Qwen and Gemma backbones. A GlobeNewswire copy dated October 19 prints Full score 63.8, second globally, public tests 62.5 also second, and quotes Sridhar Muppidi, co-founder of PurpleTalk.
Cuth and that wire copy print second globally. The lab page prints tied 3rd of 174. 63.8 is the card’s Full score. 62.52 is the public-suite run. StartLux-Decision v1.1 27B Instant’s README, dated 2026-10-16, prints Full score 69.12, first of 174 on the same 0.3.2 board. Drex Max prints public 67.59 on Nace’s own 0.3.2 bar. Those three prints stay on their own pages. Other open checkpoints are on decision models like Jev.
This site's reading
Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.
Verify
Cuth, @ItsCuthulhu, October 20, 2026, 01:50 UTC, status 2112360698403520842. The text names Sthānika in Hyderabad, Sieve 31B at 63.8 on the index's full suite, second globally, 2B, 4B, and 9B among the top of their size class, and a LoRA adapter on Gemma 4 31B with open weights on Hugging Face. 丰收, @ai_fengshou, 01:24 UTC, status 2112354110464786843, prints 63.8 and second globally. The share image on this page is the still from that post.
The Hugging Face card sthanika-ai/Sieve-31B, opened October 20. License apache-2.0. LoRA rank 128 on google/gemma-4-31B-it, revision 842da37, 979M trainable parameters, all 60 text layers. Question types choice, noul, and score. Decision Index Full score 63.8 on edition 0.3.2, board data of 2026-10-14. The authors' own run of the edition 0.3 public suite, official kit at 62d2f51, public index 62.52, raw 71.08, breadth 61.55, 140,178 scoreable requests over 43 benchmarks, 37 averaged. Area skill times 100 on that public run: Knowledge and Reasoning 49.9, Language Understanding 65.6, Retrieval and Classification 69.5, Tools and Automation 78.7, Arts and Human Taste 43.5. Latency on one RTX PRO 6000, 400 random public 0.3 requests: median 163 ms. The card says the leaderboard lists a median of 253 ms on the same GPU class. Temperature fixed at 8.5495. Token cap 32,768 per prompt row. Up to 255 options. Training: 2x B200, about 12 hours, 190,314 questions. Code github.com/sthanika-ai/Sieve. We did not load the adapter.
sthanika.ai, opened the same day, prints Sieve-31B at 63.81, tied 3rd of 174 models, against 57.62 for the same base model read the same way, and 97.8% on BFCL. Sieve-27B is 60.17, tied 5th of 149. The family is eight open-weight models from 0.8B to 31B.
A GlobeNewswire copy dated October 19, 2026 prints Full score 63.8, second globally, public tests 62.5 also second, and quotes Sridhar Muppidi. We did not rerun the suite.
Compare
63.8 is the Hugging Face card's Full score on Decision Index 0.3.2, board data of October 14. Full score weights public benchmarks 20%, private tests of the same skills 50%, and new-domain private tasks 30%. 62.52 is the authors' public-suite run. Do not merge 63.8 with 62.52.
Cuth and the GlobeNewswire copy print second globally. sthanika.ai prints 63.81, tied 3rd of 174. Those two ranks are different pages.
StartLux-Decision v1.1 27B Instant's README, dated 2026-10-16, prints Full score 69.12, first of 174 on the same 0.3.2 board. That 69.12 stays on the StartLux page.
Drex Max's product page prints public Decision Index 0.3.2 at 67.59. That bar is public benchmarks only. Do not merge 67.59 with Full 63.8.
163 ms is the card's own median on one RTX PRO 6000. 253 ms is the leaderboard median the card cites. Basal-omni-1.0-ultra's JevBench Capability 82.5 is a different board.
Terms
- Sieve-31B
- Sthānika's LoRA adapter, rank 128, on google/gemma-4-31B-it. Hugging Face sthanika-ai/Sieve-31B, license apache-2.0. Serves choice, noul, and score. The card prints Decision Index 0.3.2 Full score 63.8.
- 63.8
- Sieve-31B Full score on Decision Index edition 0.3.2, board data of 2026-10-14, on the Hugging Face card opened October 20, 2026. The authors' public-suite run on that card is 62.52.
- option-fork
- Sthānika's scoring design that scores each option as its own branch. The lab page says option order cannot change the answer under that design. Sieve-31B's card describes an answer-slot readout at a fixed temperature.