Nace AI Open-Sources Drex 1.5: A 9B Decision Model That Scores Options Instead of Generating Text

A Different Approach to Large Models: Scoring Instead of Generating

Most large language models work by generating text — you give them a prompt, and they predict and output tokens one by one. But many real-world scenarios don't actually need the model to "write" anything. Instead, they need it to make a judgment from a set of candidates: Which option carries the lowest risk? Which decision yields the highest return? That's exactly the problem Nace.AI's newly open-sourced Drex 1.5 is designed to solve.

Drex 1.5 is a 9-billion-parameter (9B) decision model whose core characteristic is: it returns a probability value for each candidate option rather than generating natural language text. In other words, its output is a quantified score for each option — not a written description that needs further parsing. This design makes it more direct and controllable for decision-making tasks.

Scoring All Options in a Single Forward Pass

Another notable engineering feature of Drex 1.5 is its ability to return probabilities for all options in a single forward pass. This stands in sharp contrast to the conventional approach — using a generative model for multi-option evaluation typically requires running inference separately for each option, or asking the model to compare them one by one in natural language, which significantly increases compute overhead and latency.

A single forward pass means that no matter how many candidate options there are, the scoring inference cost is compressed into one computation. For production environments that require real-time decision-making or high-throughput scoring, this efficiency advantage is highly practical — especially in ranking, risk control, and recommendation candidate filtering.

A forward pass is the fundamental unit of neural network inference: input data flows layer by layer through weight matrix operations until the network produces a final prediction. For generative language models, generating each token requires one forward pass, so producing a complete N-token response requires N sequential computations. Evaluating K candidate options with a generative model typically means running full inference for each option separately, with total cost scaling linearly with the number of options.

Drex 1.5 takes a different approach: it encodes all candidate options simultaneously into the input, and through its network architecture, directly outputs scoring probabilities for each option at the final layer — all in a single forward pass. At the computation graph level, this is equivalent to merging multiple sequential inferences into one parallel computation, which is especially significant for latency-sensitive real-time systems such as online recommendation and real-time risk control.

Performance and Long-Context Capability

According to officially disclosed data, Drex 1.5 achieves a score of 58.08 on the Decision Index 0.3.1 benchmark. This is a metric specifically designed to measure decision-making ability — unlike common language understanding or generation benchmarks, it focuses more on a model's performance in structured judgment tasks.

In addition, Drex 1.5 supports a context window of up to 128K tokens. Long context is especially important for decision models — real-world decisions often require synthesizing large amounts of background information, historical records, and multi-faceted constraints. A 128K window allows the model to incorporate more complete decision-relevant information in one go, without relying on repeated truncation or external retrieval stitching.

Decision Index is an evaluation benchmark designed specifically for decision-type models. Unlike general-purpose benchmarks such as MMLU or HellaSwag — which focus on language understanding or commonsense reasoning — it targets model performance in structured judgment scenarios, such as multi-attribute option comparison, risk probability estimation, and ranking consistency. Version 0.3.1 indicates the benchmark is still being iterated, meaning its evaluation dimensions and weights may evolve across versions.

As a result, the score of 58.08 currently lacks a robust cross-model reference point — without many other models' public scores on the same benchmark at the same version, it's difficult to gauge where this score stands. For production scenarios with high decision accuracy requirements, it's advisable to treat the Decision Index score as a preliminary reference and supplement it with custom offline evaluations tailored to your own business data distribution.

What Open-Sourcing Means

Nace.AI's decision to open-source Drex 1.5 gives developers and researchers a directly usable, fine-tunable decision model base. Compared to closed-source decision APIs, open-sourcing allows users to customize scoring logic for their own business needs, continue training on private data, and have stronger controllability and auditability over model behavior.

From an industry perspective, "decision models" as a direction distinct from "generative models" is being more explicitly defined. Generative models excel at creation and conversation, while decision models focus on quantified judgment and option ranking. The two are complementary at the application layer: generative models produce candidate solutions, while decision models score and filter those solutions.

Use Cases and Limitations

Drex 1.5's design is naturally suited for tasks like: risk assessment, solution ranking, A/B candidate scoring, and tool or action selection (action selection in agent scenarios). In these settings, "give each option a probability" is more directly usable than "write an explanation."

It's worth noting that the publicly available information is still fairly limited. The authority and coverage of Decision Index 0.3.1 as a benchmark, the cross-model comparison context for the 58.08 score, and generalization performance on real business data all require further independent validation. For teams considering adopting this model, hands-on evaluation against your own tasks is recommended rather than relying solely on a single benchmark score.

Overall, Drex 1.5 represents a modeling approach worth watching: treating "decision-making" as an independent task to optimize, replacing text generation with scoring output, and balancing efficiency and information completeness through single-pass inference and long-context support. For developers focused on AI production deployment — especially in agent and automated decision-making — this is an open-source option worth experimenting with.

In agent action selection scenarios, the value of a decision model is especially direct. Typical LLM agent frameworks (such as ReAct, AutoGPT, etc.) usually have the generative model handle both "thinking" and "selecting tools/actions," and coupling the two causes selection results to be influenced by generation style while making it difficult to quantify confidence levels for each action. Introducing a dedicated decision model enables a clearer division of labor: the generative model handles context understanding and enumerating candidate actions, while the decision model scores those actions and outputs a probability distribution that can be directly used for sampling or greedy selection. This architecture makes agent decision behavior easier to audit and debug, and also facilitates triggering human intervention through threshold controls.