ENPT

Agr

Last updated:

Open—$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 on HF (Agr 61 GB, Agr-flash 0.7 GB) with a pip package, a local server on POST /v1/systemone and a DESIGN.md. TypeSafe Python and JS SDKs work against it by changing base_url. No hosted API, no SLA.

  • Capability6.3/10

    Choice, Score and Noul (2–255 options) in one pass per state, questions isolated from each other. No published ECE or latency. Self-run Decision Index 0.2.1 score of 58.15 on the public kit, above the 28 Sep board leader, but not a listed board entry [vendor].

  • Adoption3/100

    Released 3 Oct 2026. 7 likes and about 80 downloads on HF (6 Oct), on the text-classification trending page. No X launch thread, integrations or production users found yet.

Vendor claims

Read this first

  1. The 58.15 is self-scored. Command Code ran the public Decision Index kit itself. Agr is not a row on the board. The board we checked (generated 28 Sep 2026) is led by Surogate Rune 26B-A4B v3 at 57.44, so the claim would put Agr first, but that is an UnverifiedClaim until the maintainers list it.
  2. Two very different sizes. Agr is a 61 GB bf16 model (Command Code serves it on 96 GB GPUs). Agr-flash is 0.7 GB on SmolLM2-360M with a trained scoring head, and the card says it is not meant for safety decisions.
  3. No latency or calibration numbers yet. The agr eval tool reports Brier, calibration error and latency on your own data, but the card publishes none of them.

What it is

Agr is a decision model from Command Code. You send a state (text or JSON) and typed questions, and it returns a probability for every option without generating text. All questions read the same state in one forward pass, but no question can attend to another. Agr reads option scores from the model’s own output layer; Agr-flash uses a small trained head.

Type Options Output
noul yes / no probability of yes
choice 2–255 named options choice, distribution, confidence
score 2–255 ordered levels expected level, distribution, confidence

The card credits Jev’s System One API, Kev, Mark Marosi’s Decider (answer-slot layout, temperature rule) and Surogate (option-order averaging). It is not affiliated with TypeSafe.

Specs

Attribute Value
Author Command Code (CommandCodeAI)
Base Agr: Gemma 4 31B IT · Agr-flash: SmolLM2-360M
License Apache-2.0 (both bases Apache-2.0, see NOTICE.md)
Weights CommandCode/agr (61.43 GB) · CommandCode/agr-flash (0.73 GB)
Token limits Agr 16,384 state / 16,384 question / 32,768 request · Agr-flash 6,144 / 2,048 / 12,288
Runtime pip install git+https://github.com/CommandCodeAI/agr, then agr serve on /v1/systemone
Launch (HF; Agr-flash 2 Oct)
Latency Not published
Pricing Self-host (free weights)

Numbers (UnverifiedClaim)

Decision Index 0.2.1: 58.15 (skill × 100, 38 counted benchmarks), run by Command Code on the full public kit. Author-reported → UnverifiedClaim.

For context, the board file we read (index-v0.2.1.json, generated 28 Sep 2026) has Surogate Rune 26B-A4B v3 at 57.44, Decider chat on Gemma-4-31B at 57.33 and pplx-decider-v1-27b at 56.40.

Fit / anti-fit

Fit when you need: an open, large System One model you can run yourself, wired to code you already wrote for the TypeSafe SDKs.

Anti-fit when you need: a hosted API, published latency or calibration numbers, or safety decisions on the small Agr-flash.

Score working (6 Oct 2026)

  • Maturity = availability 8 × 0.30 + docs 8 × 0.25 + integrations 5 × 0.25 + support 1 × 0.20 = 2.40 + 2.00 + 1.25 + 0.20 = 5.85 → 5.9
  • Capability = types 10 × 0.30 + calibration 4 × 0.25 + latency 5 × 0.25 + accuracy 5 × 0.20 = 3.00 + 1.00 + 1.25 + 1.00 = 6.25 → 6.3
  • Adoption = engagement 5 × 0.35 + ecosystem 3 × 0.35 + production 0 × 0.30 = 1.75 + 1.05 + 0 = 2.8 → 3

Accuracy stays at 5 because the score is self-run on a public kit, not a board entry. Latency 5 follows the “not measured” precedent.