Intern-Decision
InternLM's open multimodal decision models (0.8B-4B) trained on Qwen3.5, with calibration tooling and a 96-case calibration benchmark.
Overview
Intern-Decision, from InternLM (Shanghai AI Lab), turns a state, optional images and multiple typed Choice, Score and Noul questions into decisions with probabilities. It fine-tunes the language backbone of Qwen3.5 while freezing the vision tower and projector. Training uses a JSON skeleton with one placeholder per field; the logit just before each marker predicts that field's answer, so all fields share one causal forward pass, and inference restricts logits to legal answer symbols.
The release includes training code, two inference backends, temperature calibration with checkpoint presets, a browser demo, seven accuracy test suites, and a 96-case distribution calibration benchmark with generator and scorer. Training data and weights are not in the repo; weights are on Hugging Face. The README reports Intern-Decision-4B at 90.02% average accuracy across seven suites versus 88.74% for Jev, 44.16 ms mean latency on an RTX 4090, and 0.550 Brier / 0.089 ECE on the calibration pilot versus 0.595 / 0.130 for Jev, noting protocols differ. Demos show Mario, Doom and browser use.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Any typed choice over a state and optional images | — |
| Score | Ordered-level score question | — |
| Yes / no | Yes/no question | yes · no |
How it works
- Render state, images and questions into a JSON skeleton with markers
- Run one causal forward pass
- Read restricted logits before each marker and softmax
- Apply fitted temperature for calibrated probabilities
Strengths
- Open weights at 0.8B, 2B and 4B with image input
- README reports higher average accuracy and better calibration than Jev on its suites
- Fast local inference (44 ms mean on RTX 4090 for 4B)
- Ships a reproducible calibration benchmark
Caveats
- Training data and preparation pipelines are not released
- Comparisons use differing protocols, as the README notes
- Demo examples are synthetic
Getting started
Download weights from the internlm/intern-decision Hugging Face collection and follow the repository's inference backend instructions.At a glance
- Type
- Model
- Status
- research
- Language
- Python
- Runtime
- PyTorch, browser demo
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
InternLM's multimodal decision models (0.8B, 2B, 4B) fine-tuned from Qwen3.5 with the vision tower frozen: a state, optional images, and Choice, Score, and Noul questions in, probabilities out. Ships training code, two inference backends, temperature calibration, and a 96-case distribution calibration benchmark; weights on [Hugging Face](https://huggingface.co/collections/internlm/intern-decision).
AnotiaWang/awesome-decision-models
Frequently asked questions
What is Intern-Decision?
A family of open multimodal decision models that answer Choice, Score and Noul questions with probabilities.
How does it compare to Jev?
The README reports 90.02% vs 88.74% average accuracy and lower Brier and ECE, with differing protocols noted.
Are training data included?
No, only code, benchmarks and calibration presets; weights are on Hugging Face.
Related projects
Pre-registered field test of Jev, GLiNER2.5-Decide, Laya and other decision models against small LLMs on fresh September 2026 data.
ProjectresearchPythonOpen System-1 (OpenJev): Apache-2.0 research models giving calibrated per-option probabilities, best is a 4.9B exp9a.
ProjectresearchPythonLightweight reproduction of the Jev shape on a ModernBERT encoder and a diffusion LLM, comparing speed, accuracy, calibration and cost.
ProjectresearchPythonVerdict: 151M ModernBERT/GLiClass non-autoregressive decision models with calibration heads, reporting wins over Jev and Laya.
ModelresearchPythonReproducible MLX and Colab recipes that turn Qwen decoders into Jev-like typed decision models, from 0.5B on a laptop to a noncausal 27B.
ModelresearchPython- Verdict-open-jev★ 110
OpenJev: 151M ModernBERT decision model with RLCD-style calibration, a JevBench audit and an in-browser WebGPU playground.
BenchmarkresearchPython