New Multimodal Decision Models Released by InternLM

InternLM has announced the release of two new multimodal structured decision models: Intern-Decision-4B and Intern-Decision-0.8B. The models are designed to process a shared state, a schema of named questions, and optional image inputs to produce structured answers.

Technical Architecture and Inference Process

The Intern-Decision-4B model is a fine-tuned version of the Qwen3.5-4B architecture. According to documentation provided via GitHub and Hugging Face, the model utilizes a specialized inference method that differs from standard large language model (LLM) text generation.

Rather than using a traditional generate() function to produce free-form text, the model performs a single causal forward pass. During this process, the system maps question options to specific single-token symbols (such as A, B, or numeric values). The model then reads the logits at the positions immediately preceding designated placeholders in a JSON skeleton. By applying a softmax function to the allowed candidate-symbol logits and applying probability calibration, the model returns typed JSON answers.

This architecture allows the model to provide answers to every question in a single forward pass. According to discussions on Reddit, this method enables the 0.8B version of the model to achieve near-instantaneous processing speeds, even when running on a CPU.

Performance and Availability

The models are currently available through Hugging Face, with model weights, a live demo, and the underlying source code hosted on GitHub.

While official benchmark data is provided by the developers, some community observations on Reddit suggest that the 4B model may outperform existing models, such as “jev,” across various benchmarks. However, these performance claims remain unverified by independent third-party testing. Users have also expressed interest in quantized versions of the 4B model to reduce memory requirements for local deployment.

The release includes a structured API designed for candidate scoring, ensuring that the model adheres to a specific decision schema without the need for gold answers to be inserted into the prompt.