openjev

Quadrant scores

See the full quadrant

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

• Maturity5.5/10

  MIT weights, modeling helpers, SGLang package, HF Space, and a public product domain (openjev.co coming soon); still an independent research release without hosted SLA.

• Capability6.8/10

  Trained 3-way NLI enables choice via rerank and boolean-like checks; 0.8B long-context v2s plus multimodal 4B v2 (image/video/text; audio flagged as coming) broaden the surface — still not a TypeSafe choice/score/noul wire format.

• Adoption32/100

  HF likes ~379 (was ~287 on 20 Sep) plus Sam Witteveen walkthrough, game demos, and domain presence; still niche outside the open-Jev wave.

Vendor claims

• Doom from pixels ~10.4 kills/episode on v2 vs ~5.2 on v1 (author)[Vendor claim — not independently verified]
• ANLI r3 0.42→0.63 and WANLI 0.63→0.77 from v1 to v2 (author)[Vendor claim — not independently verified]
• Minecraft iron pickaxe via backward-chaining scaffold in ~22 decisions (author demo)[Vendor claim — not independently verified]
• SGLang path ~5× faster than hosted Jev (author X, 21 Sep 2026) — UnverifiedClaim[Vendor claim — not independently verified]
• RLCD calibration somehow better than vanilla GRPO (author X, 21 Sep 2026) — UnverifiedClaim[Vendor claim — not independently verified]

openjev is an independent open decision stack — Qwen3.5 fine-tuned as a cross-encoder NLI head (contradiction / entailment / neutral) and framed as a Jev-style “universal classifier” — not a TypeSafe product and not a TypeSafe /v1/systemone clone.

Weights and code: AlexWortega/openjev (MIT). Demo / long-context small checkpoint: qwen3.5-0.8b-nli-v2s-long/ (what the HF Space serves). Recommended multimodal checkpoint: qwen3.5-4b-nli-v2/ (text + images; author also describes video-text training). Also ships a text-only 4B v1, a Qwen3.5-35B-A3B MoE variant, optional latent MLP heads, and an external SGLang classify package under code/sglang_openjev/.

Product presence: openjev.co (“Coming Soon” as of 21 Sep 2026; author thanks @_akhaliq for the domain). HF Space: AlexWortega/openjev.

Typical use: score hypotheses about a shared state (predict / predict_hypotheses), rerank answer candidates, or grade against a reference. Sam Witteveen’s walkthrough (Open Jev Models Are Here!!, 20 Sep 2026) demos visual entailment on a checkout UI plus Doom/Minecraft decision loops.

Author X update (21 Sep 2026, @justALEXWORTEGA): 0.8B Qwen variant, multimodal image–video–text input with audio coming, RLCD calibration note, SGLang support claimed “5× faster than Jev” (UnverifiedClaim), and 4B / 35B + blog “this week” — treat ship dates as forward-looking, not catalog facts.

Limits (read before you ship)

• Catalog decisionTypes list choice and boolean via entailment/rerank — there is no first-class TypeSafe score / noul wire API here
• Author NLI / game numbers, RLCD vs GRPO, and SGLang “5× faster than Jev” are self-reported → UnverifiedClaim relative to ModelSystem.One benches; JevBench rankings elsewhere are third-party
• Different contract than hosted Jev or Decider System One JSON — do not drop-in swap without your own adapter
• Multimodal path needs the v2 checkpoint and enough VRAM for Qwen3.5-4B; audio is announced, not shipped as fact here

Why it is in the catalog

Public trained decision weights (not frozen-logit method-only), clear non-affiliation, and a distinct NLI/multimodal peer next to Laya / NanoJev. Roundup context: Sam Witteveen.