ENPT

MoJev

Last updated:

Open15–80ms$0/M input

Quadrant scores

See the full quadrant

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

  • Maturity6.5/10

    Academic release with preprint, public dataset, HF Space demo, full train/eval code, and MIT license. Wire-compatible API with TypeSafe SDK. Fresh release (22–23 Sep).

  • Capability7.8/10

    Multimodal (text+image), 262k native context (1M with YaRN), measured calibration ECE 0.79%, 93.23% accuracy on 12k eval. TreePacked Attention architecture documented.

  • Adoption15/100

    Day-two release with modest HF traction (14 downloads, 4 likes); academic origin may limit initial visibility vs indie/enterprise peers.

Vendor claims

MoJev is an open academic System One model from MoLeMo-Lab (appears affiliated with Fudan University) — based on Qwen3.5-0.8B with a custom TreePacked Attention scorer — wire-compatible with the TypeSafe SDK and licensed under MIT.

Pass text or image state plus runtime-defined schemas (Choice, Score, Noul); get probability distributions over permitted values in one forward pass. The server packs multiple questions into one model invocation. Native 262k context (up to 1M with YaRN).

Specs

Attribute Value
Author MoLeMo-Lab (academic, FDU-affiliated)
Base model Qwen3.5-0.8B
Parameters ~0.85B
License MIT
Launch
Status Open weights + code (GitHub, Hugging Face)
Context 262k native, 1M with YaRN
Modalities Text + Image
Decision types Choice, Score, Noul (boolean)
Calibration ECE 0.79% (measured)
Accuracy 93.23% on 12k eval decisions

Architecture

MoJev uses TreePacked Attention: each candidate attends to the state, its own question, and its own tokens. A rank-512 readout converts the encoded context/candidate pair into one logit. Training combines Plackett–Luce ranking and Brier calibration loss.

Key difference from prompt-based Jev clones: the attention mask isolates candidates while sharing state — no autoregressive decode, no text to parse.

Wire compatibility

MoJev serves a TypeSafe SDK-compatible endpoint:

pip install -e '.[transformers]'
mojev serve MoLeMo-Lab/mojev --port 8000
from typesafe_sdk import Choice, TypeSafeClient

with TypeSafeClient(api_key="local", base_url="http://127.0.0.1:8000") as client:
    result = client.system_one(
        state={"document": "I was charged twice. Please fix this ASAP."},
        questions={
            "category": Choice(
                instructions="What is this ticket about?",
                criteria={"billing": None, "technical": None, "other": None},
            )
        },
    )

Drop-in replacement for hosted Jev in local deployments.

Multimodal

Images referenced via path in state:

state = f"Identify the subject. <|vision_start|><|image_pad|><|vision_end|>{image_path}"

Visual smoke test (from README):

Candidate set Grey image P(cat) Cat image P(cat)
cat, dog 0.471 0.786
cat, dog, car, other 0.264 0.528

Limits

  • 0.85B backbone — expect held-out domains and complex reasoning to underperform larger peers like JevK5 (4B) or XOR (Qwen3.6).
  • Day-two release — verify on your domains before production.
  • Academic origin — no commercial SLA or support; MIT license allows commercial use.
  • Image support is text-interleaved — not a native vision encoder fusion like Jev-Omni.

Fit / anti-fit

Fit when you need: TypeSafe SDK drop-in for local deployment, long context (262k+), measured calibration, MIT license, academic reproducibility (preprint + dataset + code).

Anti-fit when you need: hosted inference, production SLA, larger backbone for knowledge-heavy tasks, native multimodal fusion.

Model Params Context Modalities Calibration
MoJev 0.85B 262k–1M Text+Image ECE 0.79%
Tiny-Jev 0.6B 32k Text UnverifiedClaim
XOR ~7B 260k Text+Visual UnverifiedClaim
Jev-Omni 12B ~30k T/I/A/V UnverifiedClaim

Resources