Tasks and inputs
Intern-Decision-0.8B is listed for Text classification, Structured Decisions. It accepts text, image and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Nov 19, 2026
Intern-Decision-0.8B is Shanghai AI Lab's smallest structured decision model: text and images in, calibrated choice probabilities out, in one forward pass.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Open weights under Apache 2.0 with the Qwen licence retained; no hosted price.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Some agent steps are not open questions. Which of four buttons to press, whether a request is safe, which label fits a ticket. Generating a paragraph to answer them wastes time. Intern-Decision-0.8B, from Shanghai AI Laboratory's InternLM group, answers them differently: you hand it a state, a schema of named questions with their options, and optional images, and it returns a calibrated probability for every option in every question from a single causal forward pass. No generate() call, no sampled text.
It is fine-tuned from Qwen3.5-0.8B and is the smallest of three sizes.
| Limit | Value |
|---|---|
| Questions per request | 1-16 |
| Options per question | up to 62 |
| Images | up to 8 |
| Input length | 8,192 tokens by default; longer inputs are rejected, not truncated |
| Question types | choice, score, noul (binary no/yes) |
Responses follow a Jev-compatible envelope with per-field probabilities, confidence, the chosen value and, for score questions, an expected value.
The trick is an autoregressive masked language-modelling objective. The prompt includes a complete assistant JSON skeleton with one placeholder per field. The model reads logits at the position just before each placeholder, takes a softmax over only that field's allowed option symbols, and maps them back. Then a fitted temperature calibrates the probabilities. For this checkpoint that temperature is 2.747760550703, fitted on 1,728 calibration cases; it changes confidence, not the argmax decision.
The publisher puts the seven-suite average is 79.38 for the 0.8B model, 84.68 for the 2B and 90.02 for the 4B. Mean latency on one RTX 4090 is 33.98 ms, essentially the same as the 2B model's 33.28 ms. That is the surprising part: the smallest size is not faster on that hardware. Its advantage is memory, which matters on CPUs, phones or crowded GPUs, and it scores well on structured tool decisions (94.52 on ToolACE) while struggling on harder suites (52.25 on Jevbench-Hard, 64.48 on WildJailBreak).
Use it as a gate in front of a bigger model: route confident decisions, escalate uncertain ones. The paper describes exactly that pattern with the 4B model handing low-confidence items to a larger system. Start by logging a few hundred real decisions, check that its confidence actually tracks accuracy on your data, then pick a handoff threshold. If the accuracy gap hurts, Intern-Decision-2B or Intern-Decision-4B costs little extra latency. Similar tools appear under structured decisions, including Solar Decide.
Compare the model on the work you need to complete.
Intern-Decision-0.8B is listed for Text classification, Structured Decisions. It accepts text, image and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Documented limits: 8.19K context. Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under Apache-2.0 (Qwen licence retained). Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
Microsoft
Microsoft-Decision-1.1 is a decision model for routing, triage and classification, returning calibrated probabilities. GA in Foundry from November 2026.
Liquid AI
Liquid AI d1.2 is a hosted omnimodal decision model for text, image and audio questions with calibrated answers, at $0.02 per million input tokens.
Liquid AI
Liquid AI d1.2-ultra is the top hosted decision model in Liquid's line, scoring 75 on the AA Decision Index, with audio support at $0.042 per 1M input.
Gemini Decision 1 powers Google's Decisions API, returning calibrated yes/no, choice and rubric answers instead of text, at $0.10 per 1M input tokens.
subquadratic
SubQ Snap 1.2 is Subquadratic's long-context decision model on linear sparse attention, tested to 256K tokens, priced at $0.04 per 1M input tokens.
Perplexity
pplx-decider-flash-v1.3 is Perplexity's small, half-price decision model, distilled from the 27B decider, for high-volume classification and span extraction.