InternLM ships Intern-Decision, an open-weights multimodal decision model family
InternLM (Shanghai AI Lab)
InternLM released Intern-Decision-0.8B/2B/4B on 2026-09-26, a family of structured decision models fine-tuned from Qwen3.5 that return a calibrated probability distribution over a question's options in a single forward pass instead of free-form text. The 4B variant averages 90.02 across Jevbench, Typed Decision, ToolACE, AG News and WildJailBreak, with per-checkpoint temperature calibration and ~44 ms mean latency per query on an RTX 4090. Weights are Apache-2.0 on HuggingFace with a GitHub repo and demo space.
Why it matters
Turns a small LLM into a deterministic calibrated decision engine, a cheaper primitive for agent routers and classification pipelines.
Importance: 2/5
Notable open-weights release of a specialized model family
Sources
official
internlm/Intern-Decision-4B
official
GitHub: internlm/Intern-Decision