internlm / Intern-Decision-0.8B

Intern-Decision-0.8B

Demo | Model Weights | GitHub

Intern-Decision-0.8B is a multimodal structured decision model fine-tuned from Qwen3.5-0.8B . It accepts a shared state, a schema of named questions, and optional images, and returns an answer distribution for every question in one model forward pass.

How inference works

1. Preserve the question and option order, and map each question's options to single-token symbols A , B , …, Z , a , …, z , 0 , …, 9 .
2. Render the original system prompt, state, decision schema, and a complete assistant JSON skeleton with one <decision> placeholder per field. Preserve the checkpoint's chat template and empty thinking block.
3. Run one causal Hugging Face forward pass. For the masked-next-token decision objective, read logits at the position immediately before each placeholder .
4. Take a softmax over only that field's allowed candidate-symbol logits, then apply the checkpoint's probability calibration.
5. Map symbols back to the original option values and return typed JSON answers.

This API performs structured candidate scoring. It does not call generate() or sample free-form text. A request can contain multiple fields; no gold answers are inserted into the prompt. The inference compiler uses only state , questions , and optional images .

Benchmark results

Inference latency

Measured on a single RTX 4090 with the local HF inference path. Values are per-query end-to-end latency; they are workload and hardware dependent.

Quick start

Use Python 3.12+ . Install requirements.txt in a suitable PyTorch/CUDA environment, then import DecisionEngine from the downloaded model directory:

pip install -r requirements.txt

from inference import DecisionEngine

engine = DecisionEngine(device="cuda")  # Load once; reuse for subsequent requests.
request = {
    "state": "The customer was charged twice and asks for the extra payment back.",
    "questions": {
        "team": {
            "type": "choice",
            "instructions": "Which team should handle this request?",
            "criteria": {
                "billing": "Payments and refunds",
                "delivery": "Shipping and delivery",
            },
        },
        "urgency": {
            "type": "score",
            "instructions": "Rate the priority.",
            "criteria": ["Low", "Medium", "High"],
        },
        "refund_requested": {
            "type": "noul",
            "instructions": "Is the customer asking for a refund?",
        },
    },
}
response = engine.predict(request)  # One Python dict in, one response dict out.
print(response["answers"])

predict(request) accepts one request dictionary per call and returns a JSON-serializable Jev-compatible response. It does not read request files or mutate the supplied dictionary. Reuse the engine for each subsequent request.

The engine defaults to the checkpoint next to inference.py . To load another local copy of this same model, use DecisionEngine(checkpoint="./model-copy") . Use the inference module shipped with the selected size so its default calibration matches. backend="hf" is the default and the only implemented backend. The optional request model field does not switch checkpoints; the response model identifies the weights actually loaded by this module.

Request format

{
  "state": "The customer was charged twice and asks for the extra payment back.",
  "questions": {
    "team": {
      "type": "choice",
      "instructions": "Which team should handle this request?",
      "criteria": {
        "billing": "Payments and refunds",
        "delivery": "Shipping and delivery"
      }
    },
    "urgency": {
      "type": "score",
      "instructions": "Rate the priority.",
      "criteria": ["Low", "Medium", "High"]
    },
    "refund_requested": {
      "type": "noul",
      "instructions": "Is the customer asking for a refund?"
    }
  }
}

• choice : criteria is an ordered object mapping option values to descriptions.
• score : criteria is a list (values become "0" , "1" , …) or an ordered object with finite numeric string keys.
• noul : a binary decision with options no , then yes . Optional criteria can describe these values using no / yes or false / true keys.

Supply 1–16 questions, with up to 62 options per question. Inputs exceeding DecisionEngine(max_length=8192) (default 8192 tokens) are rejected without truncation.

Images

Set the request dictionary's images list in the intended order:

request["images"] = ["images/frame-1.png", "images/frame-2.png"]
response = engine.predict(request)

The checkpoint processor handles image resizing and token expansion. Relative paths are resolved against DecisionEngine(media_root=".") (default: the working directory). Supply up to eight images; image tokens count toward the input length limit.

Response format

answers maps each field name to:

The response follows the Jev envelope: model , answers , and usage . It also includes backend , timing , and calibration as extension fields. usage.output_tokens and usage.decision_count count scored fields, not generated text tokens. confidence for a score question belongs to its most likely category; the reported expected score can lie between categories.

Calibration

The default temperature is 2.747760550703 . It was fitted separately for this checkpoint by NLL minimization on 1,728 designated calibration cases, with 1,693 separate validation cases. Test-suite labels were not used to select the temperature.

The script follows the demo's numerical sequence:

p = softmax(candidate_logits.float())
calibrated_p = softmax(log(p) / T)

This is candidate probability calibration, not a sampling temperature . It updates confidence, the noul probability, and the expected score while preserving the argmax decision. For uncalibrated candidate probabilities, use DecisionEngine(temperature=1) . A custom temperature must be finite and positive.

License and acknowledgment

Intern-Decision is derived from the Qwen3.5 series. The original Qwen license is preserved as LICENSE-QWEN . Retain the license and applicable upstream notices when redistributing. These weights were modified by decision tuning, and this release adds the structured inference wrapper and model card. We thank the Qwen team for the original models and multimodal processor.