ENPT

Mizan

Last updated:

Open55–55ms$0/M input

Quadrant scores

See the full quadrant

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

  • Maturity5.9/10

    Apache-2.0 weights with GGUF builds (1.9 GB Q4), the same request/response body as /v1/systemone and a detailed card. Solo/small team, no hosted API or SLA.

  • Capability6.7/10

    Persian-first typed decisions on text and JSON; v3 decides from layer 20 of Qwen3.5-4B (~3.2B kept). Author: 82.5% on 3,000 Persian records (Clef-Flash 9B 83.0%), best calibrated in its test; ~55 ms on an entry GPU. Not on any board.

  • Adoption1/100

    ~30 downloads and 10 likes on the v3 repo; a Persian benchmark (Tasmim-Bench) from the same community. No integrations found.

Vendor claims

  • Mizan v3 3B: 82.5% on 3,000 real Persian records vs Clef-Flash 9B 83.0% and Clef 27B 83.4%, best calibrated; 91.1 on Tasmim-Bench, ahead of Clef and Jev; about 55 ms per decision on an entry-level GPU[Vendor claim — not independently verified]

Naming note. Our radar found this as xmanii/Mizan-1.5-4B. That URL now redirects to Nimruz/Mizan-v3-3B; Mizan 1.0 and 1.5 are kept as the v1 and v1.5 tags of that repo. The author’s 1.5-era numbers (82.0% mean on Persian tasks, ECE 0.064) are superseded by the v3 card.

Mizan is a Persian-first decision model family: a record (text or JSON) plus typed questions in, a calibrated probability per answer out, in one pass, with the same body as /v1/systemone. v3 3B is text-only and reads from layer 20 of Qwen3.5-4B’s 32. v2 4B adds images (receipts, product photos). Both have GGUF builds. Questions can be in Persian or English. All numbers are the author’s (UnverifiedClaim).

Score working (13 Oct 2026)

  • Maturity = 9 × 0.30 + 7 × 0.25 + 4 × 0.25 + 2 × 0.20 = 2.70 + 1.75 + 1.00 + 0.40 = 5.85 → 5.9
  • Capability = 9 × 0.30 + 7 × 0.25 + 5 × 0.25 + 5 × 0.20 = 2.70 + 1.75 + 1.25 + 1.00 = 6.70 → 6.7
  • Adoption = 2 × 0.35 + 2 × 0.35 + 0 × 0.30 = 0.70 + 0.70 + 0 = 1.40 → 1

Accuracy 5 and latency 5: author-run, on the author’s own Persian set, one latency figure.