ENPT

Phocinae-Largha-150M-v1

Last updated:

Open21–21ms$0/M input

Quadrant scores

See the full quadrant

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

  • Maturity5.6/10

    Apache-2.0, 144.3M on mmBERT-small. Detailed card and deployment guide; official local Python server exposes /v1/systemone, plus an MCP bridge. First-party tools, no third-party integrations or SLA.

  • Capability6.2/10

    Official JevBench (17 Oct): Intelligence 3.6, calibration 56.9, p50 0.443 s. 512-token decision window. ECE 0.017 needs an opt-in column (default 0.252); vendor GPU p50 21 ms.

  • Adoption2/100

    HF has 231 likes / 167 downloads last month; X post has 41 likes / 2,489 views. Official server/MCP are first-party; no third-party integrations or production users found.

Vendor claims

  • Publisher-reported typed-decisions accuracy 0.906 (en) / 0.848 (zh), fitted evaluation with disclosed training overlap; JevBench public 126/231 (54.5%). At τ=0.6, the publisher reports 55% fewer LLM calls and 0.9936 accuracy on the kept subset, not on the full routed set.[Vendor claim — not independently verified]

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. Small, short-context and local. The decision sequence is capped at 512 tokens. It is meant to answer the easy cases and escalate the rest to a larger model.
  2. Do not read 0.906 as a clean holdout result. The typed-decisions score is fitted, and the 10 Oct card disclosure reports training-data overlap with test rows. The held-out JevBench public result is 126/231 (54.5%), below the 58.4% gate.
  3. Calibration and routing need their conditions. Shipped ECE is 0.252; the optional calibration column lowers ECE to 0.0168 but changes Brier. The τ=0.6 headline is 55% fewer LLM calls and 0.9936 on kept cases; full routed accuracy using measured Jev escalation cases is 0.8135.

Serving and evaluation notes (10 Oct 2026)

The official phocinae-server is a first-party local Python service with POST /v1/systemone; the deployment guide also documents a separate MCP bridge. These tools update maturity; they are not evidence of third-party uptake.

The model card says the 0.906 English typed-decisions result is fitted on the dataset’s train split and discloses a 500-item error-correction pool mined partly from typed-decisions test rows (151 decisions), plus small samples from Banking77 and CLINC test rows. The 0.848 Chinese result is in-mix and uses machine-translated cases. The separate JevBench public-231 result is held out: 126/231 (54.5%), below the stated 58.4% gate.

The routing figure uses a confidence threshold swept on the evaluation set. At τ=0.6, the publisher reports 55% fewer external calls and 0.9936 accuracy on the retained subset; the full routed accuracy is 0.8135 when the escalated cases use the measured Jev result. Recheck the threshold by domain before applying it elsewhere.

The shipped calibration column has ECE 0.2519 (English); the optional bundled calibration column lowers it to 0.0168 but increases Brier from 0.0207 to 0.1218. These remain publisher-reported measurements.

Score working (10 Oct 2026; proposed update)

  • Maturity = 8 × 0.30 + 7 × 0.25 + 5 × 0.25 + 1 × 0.20 = 2.40 + 1.75 + 1.25 + 0.20 = 5.60 → 5.6
  • Capability = 9 × 0.30 + 7 × 0.25 + 6 × 0.25 + 5 × 0.20 = 2.70 + 1.75 + 1.50 + 1.00 = 6.95 → 7.0
  • Adoption = 6 × 0.35 + 1 × 0.35 + 0 × 0.30 = 2.10 + 0.35 + 0 = 2.45 → 2

Integration 5: official local server with /v1/systemone and MCP bridge, without third-party framework adoption. Latency 6 and accuracy 5: publisher measurement only, and JevBench remains below its gate.

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
JevBench open weights (live 17 Oct) Phocinae-Largha-150M-v1 Capability 30.2; Intelligence 3.6; calibration 56.9; p50 0.443 s; $0.0014/1k

Score update (audit, 17 Oct 2026). Capability 7.0 → 6.2 = types × 0.30 + calibration × 0.25 + latency × 0.25 + accuracy × 0.20 = 9 × 0.30 + 7 × 0.25 + 6 × 0.25 + 1 × 0.20 = 6.15 → 6.2. Types, calibration and latency come from the earlier working (9 / 7 / 6); only accuracy is re-derived from the board (1; was 5): JevBench Intelligence 3.6 ÷ 63.6 × 8 = 0.5, half-up. Maturity and adoption unchanged.

Hugging Face on 17 Oct (Phocinae/Phocinae-Largha-150M-v1): 664 downloads, 236 likes, licence tag apache-2.0, last modified 2026-10-15.