SSyncDev

Intern-Decision-2B

SyncDev model directory · Profile updated Nov 19, 2026

Compare with another model

Intern-Decision-2B is Shanghai AI Lab's mid-size multimodal decision model that returns calibrated choice probabilities in about 33 ms on an RTX 4090.

Key specifications

Specification score
13.2
Context window
8.19K
License
Apache-2.0 (Qwen licence retained)

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.

Inputs and outputs

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

  • Text input
  • Image input

Produces

  • Text output

Capabilities

VisionOpen weights

What is Intern-Decision-2B?

Intern-Decision-2B sits in the middle of Shanghai AI Laboratory's decision-model family, and on paper it is the best value of the three. It matches the 0.8B model's latency on an RTX 4090 (33.28 ms mean against 33.98 ms) while scoring five points higher on the publisher's seven-suite average, 84.68 against 79.38. The 4B model is more accurate still, at 90.02, but takes about 44 ms.

It is fine-tuned from Qwen3.5-2B.

The idea in one paragraph

Agents often face finite choices. Instead of asking a chat model to write out its answer, Intern-Decision fills a JSON skeleton with one placeholder per question, reads the logits right before each placeholder, and turns them into a probability over that question's allowed options. All questions are answered in one causal pass. A post-hoc temperature (2.100509348278 for this checkpoint) calibrates the probabilities without changing which option wins.

Where 2B lands on the publisher's benchmarks

Suite 0.8B 2B 4B
Jevbench-Hard 52.25 63.96 73.87
Typed Decision 77.35 79.35 80.55
ToolACE 94.52 96.45 96.45
WildJailBreak 64.48 78.33 89.86
Average 79.38 84.68 90.02
ECE (lower is better) 0.066 0.100 0.065

Two things stand out. On tool-call decisions the 2B model already ties the 4B. On safety classification and hard reasoning, it is still well behind. Note too that its calibration error is the worst of the three, so check confidence thresholds before trusting them.

Request rules

  • Up to 16 questions per call, each with up to 62 options, typed as choice, score or noul.
  • Up to eight images, which count toward the 8,192-token default limit.
  • Over-length inputs are rejected outright.
  • Use the inference module shipped with this size; its calibration default is specific to the checkpoint.

A practical deployment pattern

Put it where an agent picks among tools, UI elements or ticket categories. Send everything through it, accept decisions above a confidence cut-off, and escalate the rest to a larger model. To set the cut-off, replay a labelled sample from your logs and plot accuracy against confidence; because this checkpoint's calibration is looser, you may need a higher threshold than with the 4B.

If safety gating is the main use, the jailbreak gap argues for Intern-Decision-4B. If memory is tight, Intern-Decision-0.8B loses accuracy but not speed. For larger reasoning behind the gate, Shanghai AI Lab's Intern-S2-397B is one option. Browse alternatives under structured decisions.

How to evaluate Intern-Decision-2B

Compare the model on the work you need to complete.

Tasks and inputs

Intern-Decision-2B 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.

Capacity and cost

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.

Choosing a deployment

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.

Frequently asked questions

Qwen3.5-2B.

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

Microsoft-Decision-1.1 is a decision model for routing, triage and classification, returning calibrated probabilities. GA in Foundry from November 2026.

— context

Liquid AI

d1.2

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.

65.54K context

Liquid AI

d1.2-ultra

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.

65.54K context

Google

Gemini Decision 1

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.

262.14K context

subquadratic

SubQ Snap 1.2

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.

— context

Perplexity

pplx-decider-flash-v1.3

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.

— context