Ask a chat model to decide something and you get prose plus a confidence number nobody fitted. Intern-Decision is built the other way round: it is a decision model, not a generator. A request carries a state, optional images, and several typed questions — choice, score, and noul (yes/no) — and the model returns an answer per field with a probability attached. The repository's framing is explicit about what it does not ship: training data, the private calibration and validation records, the images, and the data preparation pipelines all stay in-house; what is published is the runtime schema, both inference backends, the benchmark scoring, the temperature calibration machinery, and a browser demo.

The mechanism is a trick with markers rather than a new architecture. The assistant input is a complete JSON skeleton with one <decision> token per field, and the ground-truth answer symbols appear only in the labels, never in the input — so ordinary causal next-token alignment makes the logit immediately before a marker predict that field's answer, and all fields are resolved in one causal forward pass. Training cross-entropy runs over the full vocabulary; at inference the logits are restricted to the legal single-token answer symbols, softmaxed, and mapped back to the original labels. Training fine-tunes the language backbone of Qwen3.5 while freezing the vision tower and projector, via XTuner at a pinned revision; the training config reads its dimensions from the base checkpoint and supports the 0.8B, 2B, 4B and 9B dense Qwen3.5 variants, of which 0.8B, 2B and 4B were released on 26 September 2026. No paper accompanies the release.

Intern-Decision-4B reaches 90.02% average accuracy across seven suites (10,751 rows, 12,351 decisions: Jevbench Easy / Original / Hard, AG News, ToolACE, Typed Decision, WildJailBreak) against 88.74% for the Jev baseline, at 44.16 ms mean request latency on one RTX 4090 in BF16 versus Jev's 109.70 ms. The 0.8B and 2B checkpoints land at 79.38% and 84.68%. The honest part of the table is where it loses: on WildJailBreak the 4B scores 89.86 against Jev's 96.29, and its Jevbench-Hard 73.87 only ties JevK5. Calibration is handled as a separate, auditable step — one positive scalar per checkpoint, fitted by NLL on a disjoint calibration split, with a replay pass that asserts zero changed decisions and artifacts bound to checkpoint hashes.

The release also bundles a 96-case known-distribution calibration benchmark — six categories, 24 families, 48 paired settings, evaluation-only, with exact reference distributions stored apart from model inputs — which scores expected multiclass Brier and ECE against those references rather than against sampled labels. Calibration moves the 4B from 0.628 Brier / 0.213 ECE to 0.550 / 0.089, below Jev's 0.595 / 0.130. The repository immediately qualifies it: the paired option variants make this "a small diagnostic study, not 96 independent trials", NLL-optimal temperature need not minimize ECE elsewhere, and differing kernels plus BF16 rounding mean the HF and XTuner backends are not claimed to be bitwise equivalent. Apache-2.0; ~721 GitHub stars and 29 forks at filing.

Model Details

Architecture DENSE
License Apache-2.0
Base model qwen3.5
multimodalopen-weightagentsefficiency

Related