Intern-Decision-4B
Shanghai AI Laboratory's multimodal decision model. A fine-tune of Qwen3.5-4B's language backbone (vision tower frozen) that answers up to 16 Choice, Score and Noul questions about a state and up to eight images in one forward pass, with a fitted calibration temperature.
Each question's options become single-token symbols and the answer is read from the logits just before a decision placeholder in a JSON skeleton; nothing is sampled. Choice takes up to 62 options and inputs up to 8,192 tokens (longer ones are rejected, not truncated). InternLM reports 80.6% on the LocalLLaMA/typed-decisions test split, 100 / 98.6 / 73.9% on JevBench's public easy, original and hard tiers (201 of 231), a 90.0% mean over seven suites (its own run of Jev: 88.7%), and ECE 0.065 on the hard tier. Siblings: Intern-Decision-2B (84.7% seven-suite mean) and 0.8B (79.4%). Weights are Apache-2.0 with the Qwen licence notices kept; the training data is not released. The third-party Decision Index 0.2.1 puts the 4B at 37.81.
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one calibrated probability
- classify — assigns a category from a fixed taxonomy
- route — sends the input to one of several destinations
At a glance
| Parameters | 4.5B |
| Base model | qwen/qwen3.5-4b |
| Maker | InternLM (Shanghai AI Laboratory) |
| Released | 2026-09-26 |
| License | apache-2.0 |
| Reported accuracy | 80.6% |
| Reported latency | 44.0 ms p50 / 44.6 ms p95 per three-question request (289 tokens) on one RTX 4090, BF16, Hugging Face path |
Get the weights
pip install systemonemodels
systemone pull internlm/intern-decision
The files are served from the maker's Hugging Face repository, internlm/Intern-Decision-4B, and verified against the checksums recorded here.