ENPT

Sieve

Last updated:

Beta19–331ms$0/M input

Quadrant scores

See the full quadrant

Scored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.

  • Maturity5.4/10

    Apache-2.0 LoRA adapters + pointer heads on Qwen3.5 (2B to 27B, plus Plus variants), code and training data on GitHub, pip install from git (sieve 0.2.x). Own loader and server, no stock Transformers path. Small team, no SLA.

  • Capability7.9/10

    Official boards (17 Oct): Decision Index 0.3.2 Full 63.81 (#7). Choice (up to 255 options), Noul, Score; per-category temperatures. Latency figures are the author's.

  • Adoption2/100

    Under 50 downloads per checkpoint and 1–4 likes. No launch post, integrations or production users found.

Vendor claims

News covering this model

publicado (OK do Fabricio 17/10 BRT). Board audit of 17 Oct 2026: see the “Board audit” section below; where it conflicts with older board statements on this page, it wins.

Read this first

  1. Beta family, adapters not full weights. Each repo is a LoRA (rank 32 on 9B-Plus) plus a pointer head; the Qwen3.5 backbone downloads at a pinned revision on first use.
  2. Only one independent number. Sieve 9B is on Decision Index 0.3 (44.59). Sieve-9B-Plus (8 Oct) is not on any board yet.

The family

Checkpoint Notes
Sieve-9B (29/09) Post-trained Qwen3.5-9B; text part only (7.94B kept); average of two rank-16 runs; DI 0.3 44.59
Sieve-9B-Plus (08/10) LoRA r32 + pointer head, one temperature per question category; needs sieve 0.2.0+; ~20 GB bf16; 16,384-token states
Sieve-2B · 2B-Plus · 4B · 27B Sieve 2B: DI 0.3 20.53

The design follows Kev: each question runs as its own row from the cached state, so questions never see each other, and the model cannot generate text.

Score working (9 Oct 2026, scored on Sieve 9B)

  • Maturity = 8 × 0.30 + 7 × 0.25 + 4 × 0.25 + 1 × 0.20 = 2.40 + 1.75 + 1.00 + 0.20 = 5.35 → 5.4
  • Capability = types 10 × 0.30 + calibration 7 × 0.25 + latency 6 × 0.25 + accuracy 7 × 0.20 = 3.00 + 1.75 + 1.50 + 1.40 = 7.65 → 7.7
  • Adoption = 3 × 0.35 + 3 × 0.35 + 0 × 0.30 = 1.05 + 1.05 = 2.10 → 2

Latency 6 because the numbers are self-run. Accuracy 7: a verifiable official board, but mid-table.

Sieve-31B and the board update (12 Oct 2026)

Sieve-31B (Apache-2.0) is a LoRA (r 128, 979M trainable parameters) on google/gemma-4-31B-it, read through the option-letter logits at a fixed temperature, using the decider-ai chat layout. Up to 255 options and 32,768 tokens per row. Sthanika’s numbers (UnverifiedClaim): Decision Index 0.3 public suite 62.52 (all 140,178 requests); median 163 ms on 400 public requests. It is not on the board yet. On the board’s public column 62.52 would sit 7th, not 2nd; against full scores it would be about 5th, but full scores include private tests it has not run.

The official board (regenerated 12 Oct, 144 entries) now lists Sieve 27B at 61.25, rank 6 (tie group 3), plus Sieve 9B Plus 52.06 and Sieve 4B 47.40. Newer repos also appeared: Sieve-0.8B, Sieve-2B-v2 and Sieve-9B-v2.

Score update: capability 7.7 → 7.9, now scored on Sieve 27B. Accuracy goes from 7 to 8 for a top-6 official board result (+0.20). Maturity and adoption are unchanged.

Board audit (17 Oct 2026)

publicado (OK do Fabricio 17/10 BRT). Checked against the official boards on 17 Oct 2026: Decision Index 0.3.2 (data of 16 Oct, 19:00 UTC), JevBench and ImageJevBench live. The boards are separate sources and are never averaged. Older board statements and rankings further down this page are superseded where they conflict with this section.

Board Row Result
Decision Index 0.3.2 (16 Oct 19:00 UTC) Sieve 31B Full 63.81 (#7, tie group 3); public 62.52 (#15); same-skills 60.89; new-domains 58.31
Decision Index 0.3.2 (16 Oct 19:00 UTC) Sieve 27B Full 60.17 (#17, tie group 6); public 60.04 (#23); same-skills 57.47; new-domains 54.8
Decision Index 0.3.2 (16 Oct 19:00 UTC) Sieve 9B v2 Full 50.89 (#57, tie group 13); public 53.35 (#60); same-skills 49.29; new-domains 45.38
Decision Index 0.3.2 (16 Oct 19:00 UTC) Sieve 9B Full 44.59 (#82, tie group 19); public 41.88 (#91); same-skills 42.85; new-domains 43.54
Decision Index 0.3.2 (16 Oct 19:00 UTC) Sieve 4B Full 46.11 (#76, tie group 18); public 50.06 (#69); same-skills 45.16; new-domains 40.35
Decision Index 0.3.2 (16 Oct 19:00 UTC) Sieve 2B v2 Full 30.84 (#120, tie group 31); public 37.26 (#109); same-skills 34.87; new-domains 21.77
Decision Index 0.3.2 (16 Oct 19:00 UTC) Sieve 2B Full 20.53 (#137, tie group 39); public 22.16 (#142); same-skills 23.66; new-domains 17.76
Decision Index 0.3.2 (16 Oct 19:00 UTC) Sieve 0.8B Full 23.06 (#132, tie group 37); public 32.26 (#119); same-skills 26.01; new-domains 16.07

Board-status statements on this page to recheck against the table (may be superseded):

  • “Sieve-9B-Plus (8 Oct) is not on any board yet.”
  • “It is not on the board yet.”

Score update (audit, 17 Oct 2026). Capability 7.9 → 7.9 (unchanged) = types × 0.30 + calibration × 0.25 + latency × 0.25 + accuracy × 0.20 = 10 × 0.30 + 7 × 0.25 + 6 × 0.25 + 8 × 0.20 = 7.85 → 7.9. Types, calibration and latency come from the earlier working (10 / 7 / 6); only accuracy is re-derived from the board (8; was 8): Decision Index Full 63.81 ÷ 60.11 × 8 = 8.5, half-up, capped at 8. Maturity and adoption unchanged.

Hugging Face on 17 Oct (sthanika-ai/Sieve-9B): 114 downloads, 6 likes, licence tag apache-2.0, last modified 2026-10-13.

Sieve 31B v3 (28 Oct 2026)

publicado (OK do Fabricio 28/10 BRT).

Sthanika AI published Sieve-31B-v3 on 27 Oct 2026 (Apache-2.0, fine-tune of google/gemma-4-31B-it). The Decision Index 0.4.1 added it on 28 Oct (06:03 BRT, data of 09:02 UTC).

Decision Index 0.4.1 (data 28 Oct) Full Rank
Fastino GLiDE 1.2 71.52 #1
Sieve 31B v3 71.10 #2
vLLM-SR d3-ultra 70.18 #3
Jev 1.13 (reference) 59.22 #51

Score. Unchanged at 7.9 capability. Accuracy was already at the cap of 8: half-up(8 × 71.10 ÷ 59.22) = 9.6, capped at 8, and #2 does not trigger the #1 exception. Maturity and adoption unchanged (20 downloads on the v3 repo so far).