Skip to main content

Interfaze AIDecision modelOpen source

Lev

PopularityRank 34 of 39 decision models

24out of 100new calculation

usage data only, range 10 to 37

Summary

Lev is Interfaze AI’s LoRA adapter for typed decisions over a state. You provide text, a ticket, an email, or JSON together with yes-or-no, Choice, and Score questions. Lev reads the answers from already computed logits and returns calibrated probabilities with zero output tokens. Its interface is compatible with TypeSafe’s /v1/systemone protocol.

The adapter sits on Qwen3.5-4B and is released under Apache-2.0. Separately loaded base weights complement the adapter. Lev is therefore not a standalone 4B weight checkpoint.

GraduallyBench Decision

Lev decides 49 of 60 test cases correctly. The median of the 33 tested decision models is 60.0 %.

Accuracy

81.7 %

49 of 60 test cases

Robustness

83.3 %

25 of 30 reworded variants

Unnecessary reviews

7.7 %

4 of 52 cases without a required review

Response time

1.3 s

median, 95 % under 1.8 s, local, GPU

Accuracy compared

Share of correct decisions in percent, higher is better

20 of 33 models

As of 11/10/2026. Source: GraduallyBench Decision

Accuracy by task familyThe tick marks the median of all 33 tested models.
Triage58.3 %7 / 12, median 58.3 %
Routing91.7 %11 / 12, median 83.3 %
Constraints91.7 %11 / 12, median 33.3 %
Tables75.0 %9 / 12, median 50.0 %
Evidence91.7 %11 / 12, median 75.0 %

Lev

Benchmark run
  • The exact base-model revision is not attested.
Show test details

Decisions

Local · German, English

Execution Apple Metal. Precision FP32.

130 valid outputs from 130 attempted and 130 planned test slots.

Measured duration p50 / p95 1,255 / 1,812 ms (130 cases). Local GPU. Quantile calculation not documented.

Automatic decision accuracy
81.67 %(49 / 60 planned cases)German, English
Automatic decision accuracy, Triage
58.33 %(7 / 12 planned cases)German, English
Automatic decision accuracy, Routing
91.67 %(11 / 12 planned cases)German, English
Automatic decision accuracy, Selection with constraints
91.67 %(11 / 12 planned cases)German, English
Automatic decision accuracy, Numbers and tables
75 %(9 / 12 planned cases)German, English
Automatic decision accuracy, Sufficient evidence
91.67 %(11 / 12 planned cases)German, English
Correct robustness variants
83.33 %(25 / 30 planned cases)German, English
Correct development cases
85 %(34 / 40 planned cases)German, English
Unnecessary reviews
7.69 %(4 / 52 assessable cases without a required review)German, English
Missed required reviews
12.5 %(1 / 8 assessable cases requiring review)German, English

Response times are comparable only within the same environment and form no shared speed ranking. Where used, the test references were not independently reviewed by humans. Test coverage and methods differ by model. Each metric stands on its own; there is no combined overall score across tests. Data compiled on 11/10/2026. How we test · Download all Decision and Audio results as JSON

All benchmark results

Every published result with its source, test setup and comparison models.

Show all 83 published benchmark results

Page 1 of 7

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

S1Bench: vitaminc-dev: accuracy (fraction) (Show measurement, test conditions, and source)
Source value
0.668
Metric
accuracy
Unit
ratio
Category
S1Bench: vitaminc-dev
Direction
Higher is better
Sample
3880
Source type
vendor-reported
Evaluator
Interfaze AI Lev model card
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. Total scored items=3880; Subsets=13; Harness=Nimble-pinned manifests; Checkpoint=Qwen3.5-4B LoRA adapter with provider serving runtime
Source Interfaze AI Lev model card
S1Bench: massive-en-US: accuracy (fraction) (Show measurement, test conditions, and source)
Source value
0.857
Metric
accuracy
Unit
ratio
Category
S1Bench: massive-en-US
Direction
Higher is better
Sample
3880
Source type
vendor-reported
Evaluator
Interfaze AI Lev model card
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. Total scored items=3880; Subsets=13; Harness=Nimble-pinned manifests; Checkpoint=Qwen3.5-4B LoRA adapter with provider serving runtime
Source Interfaze AI Lev model card
S1Bench: massive-de-DE: accuracy (fraction) (Show measurement, test conditions, and source)
Source value
0.823
Metric
accuracy
Unit
ratio
Category
S1Bench: massive-de-DE
Direction
Higher is better
Sample
3880
Source type
vendor-reported
Evaluator
Interfaze AI Lev model card
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. Total scored items=3880; Subsets=13; Harness=Nimble-pinned manifests; Checkpoint=Qwen3.5-4B LoRA adapter with provider serving runtime
Source Interfaze AI Lev model card
S1Bench: boolq: accuracy (fraction) (Show measurement, test conditions, and source)
Source value
0.827
Metric
accuracy
Unit
ratio
Category
S1Bench: boolq
Direction
Higher is better
Sample
3880
Source type
vendor-reported
Evaluator
Interfaze AI Lev model card
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. Total scored items=3880; Subsets=13; Harness=Nimble-pinned manifests; Checkpoint=Qwen3.5-4B LoRA adapter with provider serving runtime
Source Interfaze AI Lev model card
S1Bench: squad2: accuracy (fraction) (Show measurement, test conditions, and source)
Source value
0.813
Metric
accuracy
Unit
ratio
Category
S1Bench: squad2
Direction
Higher is better
Sample
3880
Source type
vendor-reported
Evaluator
Interfaze AI Lev model card
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. Total scored items=3880; Subsets=13; Harness=Nimble-pinned manifests; Checkpoint=Qwen3.5-4B LoRA adapter with provider serving runtime
Source Interfaze AI Lev model card
S1Bench: paws: accuracy (fraction) (Show measurement, test conditions, and source)
Source value
0.776
Metric
accuracy
Unit
ratio
Category
S1Bench: paws
Direction
Higher is better
Sample
3880
Source type
vendor-reported
Evaluator
Interfaze AI Lev model card
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. Total scored items=3880; Subsets=13; Harness=Nimble-pinned manifests; Checkpoint=Qwen3.5-4B LoRA adapter with provider serving runtime
Source Interfaze AI Lev model card
S1Bench: multinli: accuracy (fraction) (Show measurement, test conditions, and source)
Source value
0.89
Metric
accuracy
Unit
ratio
Category
S1Bench: multinli
Direction
Higher is better
Sample
3880
Source type
vendor-reported
Evaluator
Interfaze AI Lev model card
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. Total scored items=3880; Subsets=13; Harness=Nimble-pinned manifests; Checkpoint=Qwen3.5-4B LoRA adapter with provider serving runtime
Source Interfaze AI Lev model card
S1Bench: civil_comments: accuracy (fraction) (Show measurement, test conditions, and source)
Source value
0.76
Metric
accuracy
Unit
ratio
Category
S1Bench: civil_comments
Direction
Higher is better
Sample
3880
Source type
vendor-reported
Evaluator
Interfaze AI Lev model card
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. Total scored items=3880; Subsets=13; Harness=Nimble-pinned manifests; Checkpoint=Qwen3.5-4B LoRA adapter with provider serving runtime
Source Interfaze AI Lev model card
S1Bench: aegis2: accuracy (fraction) (Show measurement, test conditions, and source)
Source value
0.8
Metric
accuracy
Unit
ratio
Category
S1Bench: aegis2
Direction
Higher is better
Sample
3880
Source type
vendor-reported
Evaluator
Interfaze AI Lev model card
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. Total scored items=3880; Subsets=13; Harness=Nimble-pinned manifests; Checkpoint=Qwen3.5-4B LoRA adapter with provider serving runtime
Source Interfaze AI Lev model card
S1Bench: helpsteer2: accuracy (fraction) (Show measurement, test conditions, and source)
Source value
0.386
Metric
accuracy
Unit
ratio
Category
S1Bench: helpsteer2
Direction
Higher is better
Sample
3880
Source type
vendor-reported
Evaluator
Interfaze AI Lev model card
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. Total scored items=3880; Subsets=13; Harness=Nimble-pinned manifests; Checkpoint=Qwen3.5-4B LoRA adapter with provider serving runtime
Source Interfaze AI Lev model card
S1Bench: summeval-relevance: accuracy (fraction) (Show measurement, test conditions, and source)
Source value
0.358
Metric
accuracy
Unit
ratio
Category
S1Bench: summeval-relevance
Direction
Higher is better
Sample
3880
Source type
vendor-reported
Evaluator
Interfaze AI Lev model card
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. Total scored items=3880; Subsets=13; Harness=Nimble-pinned manifests; Checkpoint=Qwen3.5-4B LoRA adapter with provider serving runtime
Source Interfaze AI Lev model card
S1Bench: summeval-consistency: accuracy (fraction) (Show measurement, test conditions, and source)
Source value
0.271
Metric
accuracy
Unit
ratio
Category
S1Bench: summeval-consistency
Direction
Higher is better
Sample
3880
Source type
vendor-reported
Evaluator
Interfaze AI Lev model card
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. Total scored items=3880; Subsets=13; Harness=Nimble-pinned manifests; Checkpoint=Qwen3.5-4B LoRA adapter with provider serving runtime
Source Interfaze AI Lev model card

Page 1 of 7

Popularity compared

How often the model is searched for, used and mentioned, compared with every model of its category.

Popularity index

0 to 100, higher is more popular

20 of 42 models

As of November 2026 · Source: gradually.ai popularity index

Over time

Weekly values from October 18, 2026 to November 20, 2026

LevLaya Multilingual 25Kev 0.5B 24Drex v1.5 24Drex v2.0 23

Method changes in the timeline, a jump at these points is no market event:

  1. 1November 1, 2026: Value only with search or usage data, 30-day attention
  2. 2November 1, 2026: Method version 4 (common scale, shrinkage)
  3. 3November 9, 2026: New topics measurement V20 (daily quorum, short forms)

What is gaining popularity right now, across models, vendors, harnesses and AI subscriptions, is collected in All AI trends.

All specifications

Architecture

Model class
System One LoRA adapter for typed decisions, not standalone base weights
Source: Model class
Checkpoint
interfaze-ai/lev
Source: Checkpoint
Base model
Qwen/Qwen3.5-4B, adapted through a LoRA adapter
Source: Base model

Input and output

Input
Text, tickets, emails, or JSON as state
Source: Input

Watermarking and provenance

AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.

AI watermarking
Unknown
no source
Watermarking technique
Not known
no source

Licence and access

Access
Local execution on your GPU or CPU with an optional Jev-compatible server
Source: Access
Weights
Apache-2.0 adapter, base weights must be loaded separately
Source: Weights
License
Apache-2.0
Source: License
Shutdown
None announced
no source

Further details

Output schema
Noul, Choice, and Score with calibrated probabilities and no output tokens
Source: Output schema

Sources and data date

Show all 19 sources

Benchmark sources at a glance

Which independent evaluators tested this model. This is coverage, not a quality ranking.

No active result from the reviewed external evaluators can currently be attributed to this model. Other source evidence remains available below.

Without vendor claims and historical or disputed results; those stay labelled in the single values.

TypeEvidence and data date
Research statusResearch date October 4, 2026. 12 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Unresolved

serving-revision

Note

Starred held-out values predate the last serving update; after-update rows remain separate.

Not published

per-suite-denominator

Note

The result publishes only full-run counts and nine suite scores; a per-suite denominator cannot be assigned.

Not published

confidence-interval

Note

No confidence intervals are published for the selected run.

Unresolved

v2.1-annotation-provenance

Note

The repository provenance file captured here identifies dataset v2.0.0, not v2.1.0; v2.1 human-review details remain unconfirmed.

Not published

author-readout-comparison

Note

The pinned runner identifies the selected run as option_letter_logits; no same-manifest result using the Lev authors’ distinct one-token label-code argmax with two option orders averaged is captured.

Not published

Immutable Decision Index served checkpoint revision

Note

The current public index gives a model URL and, where available, a served_checkpoint label, but no immutable commit/digest for the exact evaluated weights.

Additional sourceLev model card (retrieved October 3, 2026; October 4, 2026; October 19, 2026) · Lev model card · Editorial description reviewed October 4, 2026
Additional sourceLev announcement (retrieved October 3, 2026) · Lev announcement · Editorial description reviewed October 4, 2026
Additional sourceSysOne-bench v2.1, pinned public results (93afcfe1bd450dc052b21768e2b95b20316ebb68) (retrieved October 17, 2026)
Additional sourceJev Decision Index public index.json (release-v4) (retrieved October 28, 2026)
Additional sourceModel metadata (retrieved October 17, 2026)123456789
Additional sourceOpenRouter: Popularity index (derived), CC BY 4.0
Additional sourceVercel AI Gateway: Popularity index (derived), CC BY 4.0
Additional sourceOllama: cumulative downloads for the popularity index
Additional sourceReplicate: cumulative runs for the popularity index
Additional sourceGoogle Trending Now: limited feed selection for Germany and the US
Additional sourceCreative Commons Attribution 4.0