ENPT

JJ-1

Last updated:

Open8–30ms$0/M input

Quadrant scores

See the full quadrant

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

  • Maturity4.9/10

    0.8B and 2B weights on Hugging Face (CC BY-NC 4.0, non-commercial) with BF16, Q8_0 GGUF, a CLI and a /v1/systemone server. 4B and 9B are private; a hosted service needs sign-up. Company press release, no SLA.

  • Capability7.2/10

    Choice / yes-no / score with a probability per option, per-type temperatures in the config. Vendor: 8–30 ms per request on one B200. The 79.20% headline is the private 9B on Jurin's own 88,813-question report; the open 2B scores 68.54% there. Not on any public board.

  • Adoption2/100

    Press release on PR Times and EIN Presswire (13 Oct). HF: 0.8B 67 and 2B 88 downloads, 0 likes. Small X pickup. No third-party integrations found.

Vendor claims

News covering this model

Read this first

  1. The headline number is from a model you cannot download. 79.20% overall and 97.43% on Japanese workflows are JJ-1-9B. The public repos are the 0.8B (65.49%) and 2B (68.54%). The family hub still says the linked repos are private; on 14 Oct the 0.8B and 2B were downloadable.
  2. Non-commercial weights. CC BY-NC 4.0, inherited from StartLux-Decision. The inference code is Apache-2.0.
  3. All numbers are Jurin’s (UnverifiedClaim), from its own six-area report. JJ-1 is not on the Hugging Face Decision Index or the Artificial Analysis Decision Index.

What it is

Jurin AI (Tokyo) announced JJ-1 on 13 October 2026: a decision model family for Japanese enterprise workflows (routing, urgency, escalation, which workflow runs next). Give it a state, text or JSON, and typed questions. It returns a choice distribution, the probability of yes, or a probability-weighted level on a scale, without generating text.

Four sizes share one interface: 0.8B, 2B, 4B and 9B. The 0.8B and 2B ship BF16 safetensors, a Q8_0 GGUF, a Python package (jj1) with a CLI, and an HTTP server on POST /v1/systemone. The GGUF runs on llama-server.

Lineage. The NOTICE says the weights are “modified weights based on StartLux-Decision-2B”, and the code is adapted from StartLux-Decision and Jurin-Jev-Inference. The config is Qwen3.5 (Qwen3_5ForConditionalGeneration), with up to 26 options per pass and one temperature per question type.

Vendor results (UnverifiedClaim)

From the JJ-1 report as quoted on the hub card: 88,813 questions in six areas, 84,153 of them hard-label, accuracy weighted by question count.

Model Overall English Japanese Weights
JJ-1-9B 79.20% 84.92% 75.66% private
JJ-1-4B 75.18% 82.25% 70.80% private
JJ-1-2B 68.54% 78.09% 62.63% public
JJ-1-0.8B 65.49% 78.66% 57.34% public

In the same table Jev 1.13 and Microsoft Decision-1 both score 76.50%, and StartLux 9B 78.02%. The press release says 8–30 ms per request on one B200 and up to 99% lower cost than generative LLMs. The card says Q8_0 was not benchmarked separately.

Score working (14 Oct 2026)

  • Maturity = availability 6 × 0.30 + docs 7 × 0.25 + integrations 2 × 0.25 + support 4 × 0.20 = 1.80 + 1.75 + 0.50 + 0.80 = 4.85 → 4.9. Availability 6: public weights, but non-commercial and only the two small sizes.
  • Capability = types 10 × 0.30 + calibration 6 × 0.25 + latency 6 × 0.25 + accuracy 6 × 0.20 = 3.00 + 1.50 + 1.50 + 1.20 = 7.20. Latency 6: one vendor range, no p99. Accuracy 6: a large but self-run report, and the best rows are private sizes.
  • Adoption = engagement 3 × 0.35 + ecosystem 1 × 0.35 + production 1 × 0.30 = 1.05 + 0.35 + 0.30 = 1.70 → 2.