Intern-Decision

Not yet independently verified. InternLM’s own model cards and code repository are the only first-hand sources; InternLM published no announcement, and the one outside write-up (an OrcaRouter blog post of 26 September) repeats InternLM’s figures without testing them. All benchmark and latency figures are InternLM’s own runs. The weights are Apache 2.0, but they derive from Qwen3.5 and carry Qwen’s licence alongside, which a business should review. We will update this when it can be confirmed, and remove this note.

Intern-Decision is a set of three open-weight decision models from InternLM, the open-model team behind the Intern series, placed on Hugging Face on 26 September 2026 in 0.8B, 2B and 4B sizes, with training and inference code on GitHub. They are fine-tunes of Alibaba’s Qwen3.5 models of the same sizes, with the vision encoder left frozen. A request gives a shared state (text, plus up to eight images), and a schema of named questions — pick one option, yes or no, or a score — and the model returns a probability for every allowed answer to every question in a single forward pass. It never generates free text, so it cannot answer outside the options it was given, though it can pick the wrong one. The model card’s inference engine rejects inputs over 8,192 tokens by default. This puts it in the same group as TypeSafe’s Jev and Interfaze’s Lev: models for routing, triage, moderation and agent decisions at volume, run on a business’s own hardware with no per-call charge. InternLM reports that the 4B model averages 90.02 per cent across seven test suites against 88.74 for Jev, and answers in about 44 milliseconds on one RTX 4090 against about 110 for Jev; the 0.8B model averages 79.38 per cent. Those are InternLM’s own measurements and are not stored as scores. The weights are released under Apache 2.0, with Qwen’s licence kept alongside; the GitHub repository holds the code but not the training data.