Is Zuse state of the art?Pending benchmarks.Announced Nov 4, 2026; no independent results yet.
Bineric AI's decision-model family from Norway: 4B, 9B and 27B research models return probabilities over caller-supplied options instead of generating text. A LoRA adapter and pointer head are trained on frozen Qwen bases, targeting 27 European languages. Zuse 27B has a public no-key live demo; 9B and 27B are offered through the API and commercial deployment options. Weights are not openly released, and Zuse 4B is not offered.
The details so far
- Lab
- Bineric AI
- Announced
- Nov 4, 2026
- API
- Live
- Open weights
- Closed
- Parameters
- 4B / 9B / 27B family
- Role
- Text-only decision models: fixed options, yes/no and ordered score questions with probabilities; no free-text generation or image input
- Bases
- 4B: Qwen3.5-4B-Base; 9B: Qwen3.5-9B-Base; 27B: Qwen3.8-27B. Frozen bases with trained LoRA adapters and pointer heads; uses Kev serving stack, not a model router
- Access
- Zuse 27B public live demo needs no API key. Bineric API offers 9B and 27B; 4B is described in the paper but not offered
- Weights and license
- Zuse weights are not openly released; commercial API and deployment access. Apache-2.0 applies to the disclosed Qwen bases, not an open license for Zuse weights
- Languages
- Trained targeting 27 European languages; broader language coverage is a lab evaluation claim, not 115 training languages
- Context
- API lists 65,536 tokens per document; paper says accuracy above 32K is unmeasured and 24 GB-card deployments are limited to 8,192 tokens
- Limitations
- Paper discloses development-set selection bias, single-seed results, undisclosed training data/recipe, task-dependent calibration, residual option-order sensitivity and prompt-injection susceptibility. Site cautions against using Zuse alone for employment, credit, benefits or legal-rights decisions
- Related
- Separate from Qwen bases, TypeSafe Jev, Kev and Cloudflare Clef; 4B/9B/27B are one decision-model family
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: Zuse 27B reaches 87.5% accuracy on 20,044 questions across 27 languages, with pooled expected calibration error 0.016 versus per-source weighted 0.035. This suite guided model selection; the paper does not claim an independent held-out evaluation.
Lab claim: on reading/topic evaluations across 210 languages, Zuse 27B reaches at least 85% in 115 languages. Training targets 27 European languages and wider transfer comes from the base model.
Sources
LABBineric Zuse paper v1 and dated submission metadataarxiv.org/abs/2611.06238v1 · Nov 4, 2026LABBineric Zuse paper: architecture, availability and limitationsarxiv.org/html/2611.06238v1LABBineric Zuse product and public live demobineric.com/zuseLABBineric founder announces Zuse paperx.com/Umair_/status/2118744647845417294 · Nov 6, 2026Tracking continues until an independent board lists it.