ENPT

Jeva-4B

Last updated:

Open47–2102ms$0/M input

Quadrant scores

See the full quadrant

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

  • Maturity4.7/10

    Open weights, Apache-2.0 since 18 Oct (LICENSE and NOTICE in the repo). Python package, GitHub docs and a project page; needs an NVIDIA GPU. No hosted API, no framework integrations found, no SLA. Academic team.

  • Capability6.5/10

    Choice (with an explicit none), Noul and Score. All numbers are the authors' own; not on JevBench, Decision Index or Artificial Analysis. Noul Brier published, no ECE. Typed Decisions p50 93 ms (authors); seconds at 1,000+ options.

  • Adoption2/100

    1,246 downloads and 41 likes on Hugging Face, 101 GitHub stars, a few X summaries since 13 Oct. No integrations or production users found.

Vendor claims

publicado (OK do Fabricio 18/10 BRT).

Jeva-4B is the released model from “Jeva: System One Models from First Principles” (Qi Li and Xinchao Wang, National University of Singapore; paper dated 13 Oct 2026). It is a System One model: you send a request and typed questions, it returns probabilities, not prose. Our 13 Oct news item covers the paper. This page exists because the weights now carry a licence.

Licence changed on 18 Oct. At 07:52 BRT the Hugging Face repo got the commit “License Jeva-4B weights under Apache-2.0”: a LICENSE file, a NOTICE file and license: apache-2.0 in the card. The decision head (head.pt) and the adapter (backbone_adapter/) are Apache-2.0; the bundled Qwen3.5-4B backbone keeps its own upstream Apache-2.0 licence (backbone/LICENSE). The commit message says no weights or configuration changed. Before that, the repo had no licence and the GitHub README said model weights “retain their respective licenses”.

What it is

Attribute Value
Backbone Qwen3.5-4B, shipped inside the repo, plus a rank-64 LoRA merged for serving and a 13.4M-parameter decision head (head.pt)
Question types Choice (optional “none of the above”), Noul (probability a statement is true), Score (expected rubric level)
Design The request and each option are encoded separately; a decision stack updates only the request state against the option vectors, so options can be cached and reordered freely
Hardware Linux x86_64 and an NVIDIA GPU; the jeva Python package pins PyTorch 2.9.1 / CUDA 12.8 kernels
Serving Jeva.from_pretrained("LIQIIIII/Jeva-4B") and model.system_one(state=..., questions=...). No /v1/systemone server of its own, and we found no support for it in vLLM or transformers serve, which read answers from a standard autoregressive model
Source Hugging Face · GitHub (Apache-2.0, 101 stars) · project page

Checked on 18 Oct 2026: 1,246 downloads, 41 likes, not gated.

Evidence (UnverifiedClaim)

Everything below is from the paper and was run by its authors on their own suite. None of it has been reproduced by a third party, and Jeva-4B is not on JevBench, the Decision Index or the Artificial Analysis index (checked 18 Oct).

  • Accuracy: on Typed Decisions it scores 78.30 against 78.70 for InternDecision-4B and 74.10 for Jev 1.13.0; the paper says it leads seven of eight slices.
  • Order: on 600 Choice examples in up to five orders it flips no decision (all-order accuracy 76.67%); Jev 1.13.0 flips 17.00% (all-order accuracy 66.67%).
  • Null: 98.66% AUROC on 3,080 matched BANKING77 requests with 16 candidates (588 with no right answer); KEV-BF16 93.61%, InternDecision-4B 84.27%, JevK5 79.44%. It catches 93.88% of absent answers at 5.02% false positives.
  • Calibration: lowest Noul Brier of the compared models (0.0184 on 600 examples). The paper does not report ECE, and says temperature scaling is a separate step.
  • Latency: Typed Decisions p50 93.1 ms (46.7 ms with shared option banks, cold starts included); long-document sets 2,102 ms uncached and 75.1 ms with every option cached. At 2,048 candidates p50 is 12 s with length-aware batching. The paper does not name the GPU in the passages we read.

Score working (18 Oct 2026, from scratch, docs/06 formulas)

  • Maturity = availability 7 × 0.30 + docs 7 × 0.25 + integrations 1 × 0.25 + support 3 × 0.20 = 2.10 + 1.75 + 0.25 + 0.60 = 4.70 → 4.7.
  • Capability = types 10 × 0.30 + calibration 4 × 0.25 + latency 6 × 0.25 + accuracy 5 × 0.20 = 3.00 + 1.00 + 1.50 + 1.00 = 6.50 → 6.5.
  • Adoption = engagement 4 × 0.35 + ecosystem 1 × 0.35 + production 0 × 0.30 = 1.40 + 0.35 + 0 = 1.75 → 2.

Availability 7: open weights with a clear licence, but no hosted API. Docs 7: model card, README with install and examples, evaluation, fine-tuning and runtime guides, project page and paper; no official SDK beyond the package. Integrations 1 and support 3: nothing in vLLM, Ollama, LiteLLM or an agent framework, and support is GitHub issues (none open) with no SLA. Types 10: choice, score and boolean (Noul). Calibration 4: a published Brier score but no ECE (the rule for a claim without ECE). Latency 6: author p50 under 300 ms on Typed Decisions, one configuration, no GPU named. Accuracy 5: public datasets but author-run, not on any board, so the proportional rule (8 × board score ÷ reference) does not apply yet. Engagement 4: 1,246 downloads, 41 likes, 101 stars, a few X summaries. Ecosystem 1, production 0.

The score will move once Jeva-4B has a row on JevBench, the Decision Index or Artificial Analysis, or an independent run.

Fit / anti-fit

Fit: research on order-invariant routing, retrieval or selection over large option sets, tasks where “none of these” is a real answer, and workloads that reuse the same options across many requests.

Anti-fit: anything that needs a hosted endpoint, a /v1/systemone server or CPU-only inference today, and any choice that rests on the author numbers alone. The model card says to keep human oversight for consequential decisions.