After more than two years of large language models (LLMs) dominating the AI world, a new class of models is rapidly emerging—Decision ModelsThey no longer generate text; instead, they specialize in one task: selecting the optimal action from a set of predefined options and providing a confidence score. This week,Amazon open-sourced its first decision model, Strands Decider 2Bwhile OpenAI also released a similar product the same week. Combined with TypeSafe’s “Jev”—the earliest proponent of this concept—a new arms race has officially begun over the question: “Should AI use a cannon to swat a mosquito?”
What Is a Decision Model? How Does It Differ from an LLM?
A conventional LLM works by “generating text”: you ask it a question, and it writes out the answer word by word. A decision model, by contrast, works by “making choices”: you provide it with a predefined set of options (e.g., “next step: query database, call API, or wait for user input”), and it directly returns one choice along with aconfidence score.
This reflects a simple engineering insight: when an AI Agent runs a workflow, many steps require no “text generation”—only “selecting the next action.” Using a hundreds-of-billions-parameter LLM to decide whether to retry is massively wasteful—slow, expensive, and prone to drifting off-task in open-ended generation.
- Output is closed: selection is restricted to predefined options; no hallucination or fabrication
- Confidence-aware: reports how certain it is about its choice, enabling robust fallback handling at higher layers
- Small and fast: typically only a few billion parameters—runs locally, with latency and cost reduced by an order of magnitude
We previously conducted an in-depth review of the category’s pioneer—Jev in-depth review, whose article explained how Jev “replaces hallucination with probability.” In short, hallucination is inherent to generative models, whereas the decision model’s closed output eliminates the hallucination pathway entirely.
Amazon Strands Decider 2B: An open-source decision model that runs locally
Amazon’s newly released Strands Decider 2B stems from a small internal AWS project. According to TechCrunch, AWS Distinguished Engineer Marc Brooker, inspired by Jev, attempted to “build his own.” This side project briefly topped the Jevbench leaderboard among same-size models—and was subsequently officially adopted, cleaned up, and open-sourced by Amazon.
Like Jev, Strands Decider 2B is a model “standing on the shoulders of LLMs”—it is built upon Qwen3.5 Its 2B version serves as the “trunk,” but the training objective shifts from text generation to outputting calibrated choices—retaining the LLM’s language understanding and commonsense reasoning while specializing in “deciding what to do next.”
For developers, it delivers several concrete benefits:
- Completely open sourceAvailable for download and use now
- Small enough to run locallyNo cloud GPU required
- Low latency and low costIdeal for embedding into high-frequency Agent loops
- Confidence-score outputsEnabling workflows to “fall back to a large model when uncertain”
Why giants collectively entered the arena after Jev
TypeSafe named the model after economist William Stanley Jevons for a reason: Jevons’ paradox states that a decrease in the cost of a resource spurs an exponential increase in its consumption. In AI terms—when the marginal cost of “intelligence” drops low enough, demand for it explodes. Decision models precisely offload frequent, repetitive “micro-decisions” from expensive large models onto cheaper, specialized models.
Since TypeSafe launched Jev, researchers have already replicated dozens of similar models. OpenAI this week followed suit with a comparable release (reportedly to cut costs for its “swarm-style Agents”). Brooker admits the real challenge lies inPushing decision speed and calibration accuracy to their limits—without sacrificing general intelligence—a delicate balancing act.
Yet he also believes this market may not be dominated solely by giants. Because decision models cost far less to train than large models, building something compelling may require only a few hundred to a few thousand dollars—well within reach for small and midsize teams. TypeSafe CEO Diogo Almeida is even more direct: “Right now, many of these efforts resemble ML engineers trying to replicate a cool architecture—not genuinely striving to make intelligence practically useful.”
What decision models mean for developers
If you’re building an AI Agent, the most practical use case for decision models isWorkflow routing—offloading judgments like “which tool to call next,” “whether to retry this task,” or “whether the result is trustworthy” from the main model to the decision model. The main model handles only tasks requiring genuine “deep thinking,” while the decision model handles high-frequency, low-cost “path selection.”
This aligns with our earlier coverage of NVIDIA Out-of-Control Agent Governance Platform Same underlying philosophy—The more widespread Agents become, the more we need a lightweight, reliable “control layer.” Decision models are the core component within this control layer. To explore more infrastructure-grade AI tools like these, visit our AI Model Library and Tool Comparison Engine.
Frequently Asked Questions (FAQ)
How do decision models differ from large language models?
Large language models generate text; decision models select from predefined options and output calibrated confidence scores. Decision model outputs are closed-set and probabilistically calibrated, making hallucinations nearly nonexistent—plus they’re smaller in parameter count, faster, and cheaper, ideal for embedding into high-frequency AI Agent workflows.
Is Amazon Strands Decider 2B free?
Fully free and open-source—download and use it today. With only 2 billion parameters and built on Qwen3.5-2B, it’s small enough to run locally on your personal computer, requiring no cloud GPU or API fees.
What is Jev?
Jev is TypeSafe’s first decision model, named after economist William Stanley Jevons. It’s the pioneering model of its kind, and this site features an in-depth Jev in-depth reviewexplaining how it replaces hallucination with probability.
Can decision models replace GPT,Claude and other large models?
No—they’re complementary. Decision models excel at high-frequency “micro-decisions,” while deep understanding, writing, and reasoning still rely on large models. In practice, engineers typically deploy decision models as routers, falling back to Claude or GPT large models like these when uncertainty arises.
Can ordinary users benefit from decision models?
Decision models primarily target developers building AI Agents. End users don’t interact with them directly—but they’ll experience improved AI product performance indirectly, via “cheaper, faster, and more stable agents.”
Want to Discover More Useful AI Tools? Explore Our AI Model Library and Tool Comparison Engineor continue reading:Jev in-depth review · GLM 5.2 Benchmark · DeepSeek V4 Pro benchmark.