
Strands Decider v2 31B is a decision model from the Strands Decider team in Strands Labs, built for agentic AI. It is a full fine-tune of Google's gemma-4-31B-it that takes a state plus typed questions and returns a probability for each option in one pass, with no generated text. It answers choice questions with 2 to 255 options, yes/no questions and score questions on a 2 to 10 scale, over a 32,768 token context. Weights are open, 63 GB in bf16 on one 80 GB GPU, and it scores 71.45 on the Decision Index 0.3.
A workable 31.27B-parameter dense decision model from Strands Agents. Treat the modality benchmarks above as the leading indicator of fit — composite scoring across modalities is still maturing.
Generated from this model’s benchmarks and ranking signals. Editor reviews refine it over time.
Access model weights, configuration files, and documentation.
Benchmark data sourced from: Hugging Face, Arena.ai, Artificial Analysis.
The top devices for this model at 4-bit, ranked by fit and speed.
| Device | Grade | VRAM |
|---|---|---|
| Acer Veriton GN100 AI MiniAcer | SS | 26.8 GB |
| AI TOP ATOM ATAGB10-9003Gigabyte | SS | 26.8 GB |
| AMD Instinct MI300XAMD | SS | 26.8 GB |
| AMD Instinct MI325XAMD | SS | 26.8 GB |
| AMD Instinct MI350PAMD | SS | 26.8 GB |
Cheapest current cloud rentals with at least 27 GB VRAM, refreshed hourly.
Advertising disclosure: we earn commissions when you shop through the links below.
| Option | Cost / GPU-hour |
|---|---|
NVIDIA RTX PRO 5000 BlackwellVast.ai · Spot · 48 GB VRAM | $0.16 |
NVIDIA A100 80GB PCIeVast.ai · Spot · 80 GB VRAM | $0.20 |
Strands Decider v2 31B is a dedicated decision model developed by the Strands Decider team at Strands Labs, specifically engineered for routing, triage, and discrete choice execution in agentic AI architectures. Built as a full fine-tune of Google's gemma-4-31B-it, the model moves away from autoregressive text generation. Instead of predicting the next token in a conversational sequence, it takes a system state alongside typed questions and executes a single forward pass to output calibrated probabilities across predefined options.
At 31.27 billion parameters, the model establishes a distinct middle ground between lightweight edge classifiers and compute-heavy generative foundation models. Scoring 71.45 on the Decision Index 0.3 benchmark, it ranks among the top open-weights decision engines available. For developers looking to run Strands Decider v2 31B locally, it eliminates the unpredictable latency, JSON schema parsing failures, and hallucination risks common when using standard generative large language models for tool selection and deterministic routing.
By packaging decision logic into a dense 31.27B architecture, Strands Agents provides an enterprise-grade "system one" model for autonomous pipelines. It evaluates choices across a wide 32,768-token context window without generating conversational filler, returning raw probabilities, expected scores, and reliability metrics directly to your runtime.
Strands Decider v2 31B inherits the foundational dense transformer architecture of Google's gemma-4-31B-it, maintaining a uniform parameter distribution across all layers rather than employing a Mixture of Experts (MoE) routing mechanism. Every one of its 31.27 billion parameters is active during inference, providing stable, predictable execution times across batch workloads.

Explore the Provider
Aggregate stats, leaderboard, release timeline, and benchmark coverage across every Strands Agents model we track.

Check which of your devices can run this model and how fast.

Compare hosted per-token prices before you commit to local hardware.

Break-even math for self-hosting, renting a GPU, or paying per token.
Live status for the hosted APIs you might use instead of running this locally.
The model deviates from standard instruction-tuned language models in its output formulation:
Strands Decider v2 31B is text-only and does not produce generative text, code, or conversational dialogue. Its inference engine is purpose-built to resolve three strictly typed decision primitives:
In real-world local AI infrastructure, these capabilities translate directly into high-efficiency agent workflows:
Deploying a local AI model with 31.27B parameters in 2026 requires understanding the balance between quantization, memory bandwidth, and single-pass inference characteristics.
Understanding how to run a 31.27B model on consumer GPU setups comes down to memory allocation:
Evaluating Strands Decider v2 31B performance differs from standard LLMs. Traditional tokens per second metrics do not apply during generation because the model does not run an autoregressive decode loop. Instead, Strands Decider v2 31B tokens per second are measured strictly during the prefill phase (context processing).
On an RTX 4090 at Q4_K_M, prompt ingestion runs at roughly 2,200 to 3,500 tokens per second depending on context length. Once the 32K state is encoded, the classification forward pass resolves in 15 to 40 milliseconds. This makes end-to-end decision latency exponentially faster than prompting an autoregressive LLM to output a structured JSON decision.
The model can be served using the official strands-decider Python package or integrated via standard local inference runtimes such as Ollama or vLLM:
1pip install strands-decider2strands-decider serve StrandsAgents/strands-decider-v2-31b --port 8080
For custom agent loops, send a structured payload over HTTP specifying the shared state and your target evaluation questions.
When evaluating Strands Decider v2 31B hardware requirements and capabilities, practitioners typically compare it against two classes of models: specialized decision engines and general-purpose 32B foundation models.
MachinePulse's karpo-decision-31B currently leads the open decision space with a 73.9 Decision Index score, compared to Decider v2's 71.45. However, Strands Decider v2 31B provides tighter integration with standard agent development frameworks and native calibration metrics that yield more predictable confidence intervals on unseen routing tasks. If raw benchmark classification accuracy is your only metric, karpo-decision holds a slight edge. For production pipelines requiring calibrated reliability scores alongside out-of-the-box serving tools, Decider v2 31B is easier to operate.
Running a general model like Qwen 2.5 32B or the base Gemma 2 27B/Gemma 4 31B with constrained JSON decoding requires substantial VRAM and suffers from autoregressive latency overhead. A standard generative model must output tokens one by one, often taking 500 to 1,500 milliseconds to complete a routing thought trace and output schema. Strands Decider v2 31B completes the same classification in tens of milliseconds with zero JSON syntax errors. The tradeoff is flexibility: Decider v2 cannot draft an email or explain its reasoning in natural language. Use a standard LLM when open-ended generation is mandatory; deploy Strands Decider v2 31B locally when you need fast, deterministic control flow inside complex agent loops.
| $0.32 |
NVIDIA RTX A6000RunPod · Community · 48 GB VRAM | $0.33 |
NVIDIA A100 80GB PCIeVast.ai · On-Demand · 80 GB VRAM | $0.37 |
Per-GPU rate across RunPod, the Vast.ai marketplace, DigitalOcean, and Vultr.
Spot tier is interruptible. Plan for restarts when comparing against on-demand prices.