inclusionAI released LemonDecider, a 4B decision model with explicit abstention
inclusionAI released LemonDecider, a 4B-parameter decision model derived from Qwen3.5-4B, on Hugging Face under Apache-2.0. It uses a 255-way single-token readout for choice, noul, and score questions with explicit abstention. The model is available for download with custom code and a standalone CLI.
China context
- Original name
- inclusionAI
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Builders outside China can download the weights and use the model for structured decision tasks, but must implement the custom prompt rendering and handle the 8192-token limit.
- For investors
- The release of a specialized decision model by an Ant Group entity signals continued investment in open-weight model variants targeting enterprise decision workflows.
LemonDecider is a typed decision engine built on Qwen3.5-4B, replacing the LM head with a 255-way single-token readout. It supports choice, noul, and score question types, with an explicit unknown candidate for abstention. The model was distilled from imajev-4b and reports a dev set accuracy of 0.7735 at temperature 1.0. It is not a generative chat model and requires left padding and a maximum input length of 8192 tokens.
The model uses a custom AutoJevForDecisionClassification architecture with a linear head of size 2560×255 over the last prompt position. It rejects KV-cache decoding and right padding. The decision semantics include renormalized probabilities over known candidates and a separate unknown probability. A temperature fit on the dev set found T=0.60 improved NLL and Brier score, but the checkpoint still serves at T=1.0 pending public evaluation.
Developers outside China can now use a specialized decision model from Ant Group's inclusionAI, reducing the need to adapt general-purpose LLMs for structured choice tasks. The Apache-2.0 license permits commercial use, but the model's custom code and prompt-layout requirements may increase integration effort.
LemonDecider offers a lightweight alternative for applications requiring structured decisions with explicit abstention, such as classification, scoring, or yes/no queries. Its 4B parameter size and single forward pass inference may reduce latency and cost compared to larger generative models.
The model card states that adopting the fitted temperature is blocked until the public evaluation suite is re-evaluated. A specific signal to watch is whether inclusionAI releases updated evaluation results or a new checkpoint with temperature 0.60.