Darwin-27B-RSI-Decision: A Single-Request Decision Engine Returning Yes/No, Choice, and Score Answers with Probabilities

TL;DR: Darwin-27B-RSI-Decision is a specialized model from VIDRAFT designed to handle three distinct decision-query types — binary (yes/no), multiple-choice, and scalar scoring — within a single inference request, returning each answer alongside calibrated probabilities. It targets developers who need structured, probabilistic decision outputs without chaining multiple model calls. The model has been mirrored on Gitcode to reach the Chinese developer community via CSDN.

What it is

Darwin-27B-RSI-Decision is a 27-billion-parameter decision-focused language model developed by VIDRAFT. Rather than producing open-ended text, it is purpose-built to answer structured decision queries and return outputs in one of three formats per request:

  • Binary decisions — yes or no answers with an associated probability score
  • Selection decisions — choosing from a provided set of options, again with per-option probabilities
  • Scalar scoring — assigning a numeric score to a subject, accompanied by a confidence or probability value

The key developer-facing proposition is that all three query types are handled in a single API call, removing the need to route different question formats to different models or post-process raw text into structured decisions yourself. The model is being distributed to Chinese developers through a Gitcode mirror, following coverage on CSDN, one of China's largest developer communities.

How it works

At a conceptual level, Darwin-27B-RSI-Decision is trained to treat decision-making as a classification and regression task over a structured output space, rather than a next-token text generation task.

When a developer submits a query, the model interprets the decision type (binary, choice, or score) from the request structure, then produces:

  1. A discrete or continuous answer in the appropriate format
  2. A probability distribution or confidence value accompanying that answer

This design reflects an approach sometimes called decision-aware fine-tuning — where a base large language model is adapted so that its output layer is aligned to produce well-calibrated probabilistic judgments rather than free-form tokens. The "RSI" designation in the model name suggests a specific training methodology developed internally at VIDRAFT, though the precise details of that methodology are not publicly disclosed.

The result is a model that behaves more like a structured inference engine than a general-purpose chatbot — suitable for use cases such as content moderation decisions, candidate ranking, risk scoring, automated triage, and recommendation filtering, where downstream systems need machine-readable, probabilistic outputs rather than natural language explanations.

The OpenAI-compatible API surface means developers can point existing client code at Darwin-27B-RSI-Decision with minimal changes, treating it as a drop-in decision layer.

Benchmarks & results

No specific public benchmark numbers or evaluation datasets are cited in the source article. The coverage focuses on distribution and usage guidance rather than comparative performance tables. Developers evaluating the model for production use should conduct their own task-specific evaluations against relevant decision benchmarks (e.g., calibration metrics such as ECE, accuracy on binary classification holdouts, or ranking metrics for the scoring mode).

VIDRAFT has not published third-party benchmark comparisons for this model in the available source material.

How to try it

The model is being distributed via a Gitcode mirror targeting Chinese developers, as announced through CSDN. For developers elsewhere, VIDRAFT models are generally accessible through the following public channels — check the official VIDRAFT Hugging Face organization and GitHub for current availability:

# Check Hugging Face for the model (replace with the exact published model ID when available) huggingface-cli download VIDRAFT/Darwin-27B-RSI-Decision 

If VIDRAFT exposes an OpenAI-compatible endpoint for this model, you can query it using standard tooling:

curl https://<published-api-endpoint>/v1/chat/completions \ -H "Authorization: Bearer $YOUR_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "Darwin-27B-RSI-Decision", "messages": [{"role": "user", "content": "<your decision query here>"}] }' 

⚠️ Note: The endpoint URL and exact model identifier above are placeholders. Use only officially published endpoints from VIDRAFT's documentation. Do not use invented or unverified URLs in production.

The Gitcode mirror is the confirmed distribution channel mentioned in the source article for Chinese developers.

FAQ

Q: Why does this model return probabilities alongside answers — isn't that just a softmax score?
A: Calibrated probability outputs are specifically trained for, not simply taken raw from a softmax layer. Well-calibrated probabilities mean that when the model says "70% yes," that confidence score should reflect actual empirical frequency — which requires specific training signal beyond standard cross-entropy loss. This makes the outputs meaningful for downstream decision thresholds and risk management logic.

Q: Can Darwin-27B-RSI-Decision handle all three query types (binary, choice, score) in a single batch request, or only one type per call?
A: According to the source, the model is designed to handle all three query types within a single request — that unified handling is the core design goal. For specific batching behavior and request schema details, refer to VIDRAFT's official API documentation.

Q: Is this model suitable for real-time production inference given its 27B parameter size?
A: At 27 billion parameters, latency will depend heavily on your hardware configuration and quantization choices. For latency-sensitive applications, evaluate quantized variants if VIDRAFT makes them available, or test throughput against your SLA requirements before committing to production deployment.

Originally reported by CSDN (중국) (2026-11-15) — source article.