SSyncDev

Intern-Decision-4B

SyncDev model directory · Profile updated Nov 18, 2026

Compare with another model

Intern-Decision-4B is Shanghai AI Lab's 4B model that answers a whole schema of multiple-choice questions with calibrated probabilities in one forward pass.

Key specifications

Specification score
9.7
License
Apache-2.0

The specification score summarizes recorded features and limits. It is not a benchmark result.

Pricing conditions: No API pricing stated; open weights for self-hosting

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-4B?

Routing a support ticket, flagging a jailbreak and choosing a tool are all the same problem underneath: pick one option from a short list, and know how sure you are. Generating free text to make that choice is slow and hard to calibrate. Intern-Decision-4B does not generate at all.

You give it a shared state, a schema of named questions with their options, and optional images. It returns a probability distribution over the options for every question at once, from a single forward pass. Shanghai AI Laboratory fine-tuned it from Qwen3.5-4B and released it on 26 September 2026.

The mechanics

  1. Options for each question are mapped to single-token symbols (A–Z, a–z, 0–9), preserving order.
  2. The prompt, state, schema and a JSON skeleton with one placeholder per field are rendered with the checkpoint's own chat template.
  3. One forward pass runs. Logits are read at the position just before each placeholder.
  4. A softmax over only that field's allowed symbols, plus the checkpoint's calibration, gives the probabilities.
  5. Symbols map back to typed JSON answers.

No generate() call, no sampling. That makes outputs deterministic and cheap.

Publisher-reported results

Model Average Brier ↓ ECE ↓
Intern-Decision-0.8B 79.38 0.530 0.066
Intern-Decision-2B 84.68 0.437 0.100
Intern-Decision-4B 90.02 0.347 0.065

The average spans Jevbench, a typed-decision set, ToolACE, AG News and WildJailBreak. On a single RTX 4090 the smaller siblings run at about 33 ms mean latency per query.

Good fits

  • Classification with several labels per item, where you want one call instead of several.
  • Guardrails that need a score, not a yes or no.
  • Tool or route selection inside an agent loop, where a large model's latency hurts.

Trying it on your data

Take a labelled sample of a few hundred real decisions. Compare accuracy and calibration against your current prompt-based classifier, and plot a reliability curve, because calibration is the selling point. Start with the 4B and drop to 2B or 0.8B only if latency demands it. Related options live in the structured decisions directory and among text classification models. The licence is Apache-2.0.

How to evaluate Intern-Decision-4B

Compare the model on the work you need to complete.

Tasks and inputs

Intern-Decision-4B 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

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. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

No. It scores candidate options with one forward pass and returns calibrated probabilities as typed JSON, without calling generate().

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