Strands drops a 2B model that ditches text to decide in 115ms

Strands Decider 2B is a tiny, 2-billion-parameter open source decision model built for one job: picking between predefined choices and scoring its own confidence. It won't chat with you or summarize documents, but it makes categorical choices absurdly fast.

Stripping the text head for pure speed

Most LLMs are designed to generate text token by token. That flexibility is great for conversational bots, but it adds massive latency when all you need is a simple yes-or-no classification.

Strands built Decider 2B by taking a Qwen3.5-2B backbone and chopping off the language modeling head entirely. In its place, they attached a custom pointer head with just over one million parameters, fine-tuning the base model using a rank-16 LoRA adapter.

The pointer head scores candidate options in a single parallel pass by comparing the hidden state of each option to an answer token. The public release is version 19 of the architecture, following an earlier iteration that used a slot head design with weaker performance.

Because it skips autoregressive decoding, execution times drop dramatically. On a standard local Nvidia RTX 3090, Decider 2B hits a median decision latency of around 115ms. Even on an M3 MacBook, small tasks run with a median latency of roughly 153ms.

A 100ms guardrail for local agents

You won't use this model to write code or draft essays. It is also considerably worse at complex multistep reasoning than frontier reasoning models. But inside an agentic pipeline, it acts as an ultra-fast control valve.

Common use cases include model routing, tool selection, context trimming, and policy checks. In one demo, Strands placed Decider 2B inside a tool-execution hook. When an overeager agent tried to fetch weather data without knowing the user's city, Decider 2B took around 100ms to flag that the input lacked grounding, stopping the call before an expensive external API fired.

On the public JevBench benchmark, Decider 2B ranked 3rd out of 33 models in the 2B category for accuracy and calibration, and 1st among models strictly under 2B parameters. It also solved 100% of the benchmark's easy tasks.

The project is fully open source. You can grab the code and CLI on GitHub, download weights from Hugging Face, or inspect the training datasets and scripts released alongside the weights.

Why it matters

If you are building autonomous agents locally and fighting high latency or bloated API bills, Decider 2B is an immediate win. Running a full LLM just to check whether an argument is valid or which tool to call next is slow and expensive.

We think this is a smart architectural pivot. By trading away open-ended generation, Strands delivered a responsive, local component with genuine confidence calibration. Decision models like this will likely become standard glue in agent pipelines, filtering noise at the edge while saving heavy frontier models for the hard thinking.