A
ANT GROUP / lemondecider-4b-20261112-v3.1

LemonDecider 4B v3.1

LemonDecider revision based on Qwen3.5-4B, published in the official repository on November 17. It processes a text or JSON state and independent choice, noul or score questions. It returns probabilities or scores instead of writing answers.

01 / SUMMARY

The essential

ORIENTATIONTyped decisions on text and JSON

Primary use stated or inferred cautiously from official documentation.

CONTEXT8,192 tokens per question, including the template; up to 255 options

Maximum input capacity when published by the source.

EXITChoice probabilities, yes/no or ordinal scores; no text generation

Documented maximum generation limit.

ACCESSHugging Face · downloadable checkpoint

Verified availability channels.

MODEL TYPEText classification

Primary functional family and verified specialties.

LICENSEApache-2.0

Declared terms for API access, model weights, or self-hosting.

02 / TECHNICAL SHEET

Limits and integration

API IDlemondecider-4b-20261112-v3.1
Model typeText classification
Access modelOpen weights
LicenseApache-2.0
DeploymentSelf-hosted
Release17 NOV 2026
Knowledge cutoffNot published
EntranceText · JSON
ExitProbabilities
Context window8,192 tokens per question, including the template; up to 255 options
maximum outputChoice probabilities, yes/no or ordinal scores; no text generation
ReasoningNot published
Published tools
Structured OutingsNot published
Batch processingNot published
Prompt cacheNot published
Fine-tuningNot published
Verified platformsHugging Face · PyTorch · Transformers
BEST SUITED FOR

Routing, classification and selection among explicit alternatives. The fused BF16 weights are self-contained; probability postprocessing uses FP32. Capacity includes the template and the complete state for each question.

WORTH MONITORING

The runtime needs repository code that should be reviewed before execution. Images and audio are unsupported. Inputs are not automatically truncated. Earlier development results are not a fresh full evaluation of the published BF16 package.

03 / CAPABILITIES

What it can do

01

Orientation

Routing, classification and selection among explicit alternatives. The fused BF16 weights are self-contained; probability postprocessing uses FP32. Capacity includes the template and the complete state for each question.

02

Context

8,192 tokens per question, including the template; up to 255 options · Choice probabilities, yes/no or ordinal scores; no text generation

03

Tools and integration

Not published

04

Access

Hugging Face · downloadable checkpoint

Input modalities
TextJSON
04 / PRICES

Documented cost

ConceptWorthUnit/condition
Standard inputNo published official API rateStandard API rate
Cached inputNot publishedReading reused prefixes
Cache write or storageNot publishedThe condition varies by provider
Standard outputSelf-hosted infrastructure costMay include reasoning tokens
Batch inputNot publishedAsynchronous processing
Batch outputNot publishedAsynchronous processing

Apache-2.0 weights, with self-hosting infrastructure costs. No official API tariff has been verified. The identifier contains 20261112, but this revision was published on November 17; the suffix is not treated as the release date.

i Prices change and may depend on level, region or context length. Check the source before making a decision.

05 / EVALUATIONS

How to read the results

“

The card distinguishes a local development evaluation from a published-runtime check on 91 requests and 444 questions. That sample validates integration with the same fused weights but does not replace a fresh full evaluation. Historical scores are not assigned to the current package.

i Historical development results; no fresh full evaluation of the published package.

Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.

06 / VERSIONS

Chronology

lemondecider-4b-20261112-v3.1

Revision 8eca78d5: full weights and independent questions; retains the published internal identifier.

07 / SOURCES

Traceability