Decider is an open-source model interface for applications that need a decision over a predefined set of answers, rather than a paragraph of generated text. Give it a state and typed questions, and it returns a selected answer with a probability distribution in one model forward pass. That can simplify routing, scoring, and other structured workflows—but confidence is not proof that an answer is correct.
What Decider returns
Decider makes the answer space explicit before inference. Its documented question types represent three different shapes of decision:
As an Amazon Associate I earn from qualifying purchases.
- Choice: select among named options, such as which team should handle a case.
- Score: choose among described levels, such as rating how frustrated a customer seems.
- Noul: estimate a yes-or-no outcome, such as whether a refund action is needed.
For each question, the software can return the selected option or level, its confidence, and the probability distribution across valid answers. The output is structured decision data intended for software to consume; it is not a general-purpose prose response. The project describes a system_one interface and an HTTP serving path. The project repository documents the interfaces and examples.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
How one forward pass produces a decision
In the described implementation, answer slots are placed alongside the questions in the model input. The model runs once; Decider reads the logits at those slots, applies a fitted temperature, and normalizes the scores over the permitted options. Because the answers are constrained in advance, the calling application does not need to ask a free-form model to generate an answer and then parse that text back into a category.
#1 Best Overall
This is a workflow advantage, not a guarantee of reliability. A probability distribution describes the model’s relative confidence across the allowed answers. The project author warns that a model can be well calibrated and still be wrong; accuracy and calibration vary with the model, task, and difficulty. Calibration measures should therefore be considered separately from whether a particular decision is correct. The author’s technical explanation discusses the method and its limits.
Where it fits—and where it does not
Good fit: bounded decisions in software
Decider is most naturally suited to tasks where the application can define a finite answer space and needs a machine-readable result: classifying a request, assigning a queue, choosing a score band, or estimating a yes/no action. It may be useful when downstream code needs probabilities as well as a selected label.
Rank #2
- 🎓 Educational & Fun for All Ages: Dive into financial freedom with "Paycheck to Billionaire," a strategic board game that imparts skills in cost management and smart investing from startup to empire. Ideal for family game nights and educational play, this game is engaging for adults and kids ages 8 and up. 🌟
- 🎲 Interactive Family Game Time: Bring excitement to your family gatherings and game nights with this engaging board game that offers a fun way to learn about economics without the complexity of bankruptcy or monopoly rules. Ideal for kids, teens, and adults. 🎉
- 🎉 Perfect Party Game: Light up the party with "Paycheck to Billionaire," where young entrepreneurs aged 8-12 and older can strategize and compete to build their business empire. It's a hit at parties, fostering lively interaction and competitive spirit. 🏆
- 🍎 Teacher’s Educational Pick: A must-have for educational environments, this game allows teachers to introduce financial concepts and critical thinking skills in a fun, interactive manner. It's an excellent tool for classrooms and homeschool settings, encouraging students to engage with real-world economic strategies. 📚
- 🎁 Perfect for Building Business Mind: "Paycheck to Billionaire" is the ideal gift for minds keen on entrepreneurship. This engaging board game offers a fun and educational pathway to understanding business and finance, making it a superb choice for nurturing future business leaders during family game nights or as a thoughtful present. 🌐
Use caution with multi-step reasoning
The author says the released approach does not handle multi-step arithmetic or multi-hop reasoning well. A complex judgment may need to be broken into smaller, explicit questions; even then, each intermediate result needs evaluation in the target workflow. The author also notes that rule-following depends on patterns present in training data, that hard-item calibration is a weaker area for the 2B model, and that released models’ text-state game results are not solved zero-shot. These are author-reported limitations, not independent evaluations.
The same write-up described the released models as English-only at publication in September 2026. Language coverage can change with later versions, so confirm the behavior of the specific model you deploy rather than assuming broader multilingual support.
Rank #3
- Package Dimensions: 5.207 cms (L) x 32.791 cms (W) x 32.994 cms (H)
- Product Type: Tabletop Game
- Package Quantity: 1
- Country Of Origin: China
Choosing a model and deployment route
Model size, precision, runtime, and hardware affect both memory needs and latency. The project’s current repository guidance gives the following approximate weight-memory requirements for the specified configurations:
| Model and configuration | Approximate memory guidance | What the figure applies to |
|---|---|---|
| decider-2b, bf16 | 4 GB | Project repository guidance for bf16 weights |
| decider-4b, bf16 | 8.4 GB | Project repository guidance for bf16 weights |
| decider-35b-a3b, bf16 | 65 GB | Project repository guidance for bf16 weights |
| decider-35b-a3b, NVFP4 with vLLM | 19.6 GB | Project repository guidance for this precision and serving setup |
These are project guidance figures for model weights and named configurations, not universal total-system requirements or a promise that a model will fit every deployment. Allow for runtime and application overhead, and check the model’s instructions for the actual setup. The repository documents CUDA inference, Apple Silicon/MPS support for some dense models, GGUF files served through llama.cpp, and CPU inference. A dedicated GPU is not required for every route; the project also documents a CPU/laptop path for the 2B model. Repository installation and deployment instructions describe the available paths.
The documented Python installation is pip install decider-ai, while the package is imported as decider. Model weights are loaded from Mapika’s model repositories. The PyPI listing identifies the package as open source, and the 2B model card specifies the Apache-2.0 license. The package page and the 2B model card provide those details.
Decider is an independent reproduction of the System One model class; it is not an official Jev or TypeSafe AI product. The repository explicitly says the project is not affiliated with or endorsed by TypeSafe AI.
Best Value
- 🎲Multifunction-Seek a combination of durability, amusement, and pedagogical values in the funnysquee's 3.15 Inches Dice Toy! This one-of-a-kind toy fits seamlessly into your game nights and enlivens mathematical teachings, ensuring a fun time for kids and adults.
- 🎲A Game Changer - Make your playtime more exciting with our dice toy! It's a perfect tool for teaching probability, simple arithmetic, or even serve as markers or counters. Let teachers stimulate students' interest and foster critical thinking via a dynamic, hands-on approach.
- 🎲Convenience at Its Best - Say goodbye to squinting and guessing, thanks to the massive and easily decipherable numbers on each face of our 6-sided dice. Their ergonomic design and comfortable grip make your gaming experience smoothly enjoyable.
- 🎲Gift of Joy - Think Birthday, Christmas, Halloween, New Year's Day, Valentine's Day, Mother's Day, Father's Day, and even Anniversaries. Our dice toy is the perfect surprise! It's ready to bring joy, facilitate educational activities, or spice up board games.
- 🎲Commitment to Satisfaction - Suitable for individual enjoyment and an ideal addition to any classroom or homeschooling setup. Should you encounter any issues, our responsive customer service team is ready to assist you within 24 hours.
What the published performance figures do—and do not—show
The available numbers come from different evaluations and conditions. They are useful context, but they are not one controlled comparison of all model sizes or deployments.
| Reported result | Model or evaluation | What to keep in mind |
|---|---|---|
| 0.784 held-out regression accuracy | decider-4b, Mapika project repository, 2026 | Repository task metric; not a Decision Index score |
| 0.810 held-out regression accuracy | decider-35b-a3b, Mapika project repository, 2026 | Repository task metric; not directly comparable to the Decision Index figures below |
| 54.3 score; fourth place | decider-35b-a3b on Decision Index edition 0.1, reported by author Mark Marosi in 2026 | The index’s reported score and rank, not a general accuracy percentage |
| 3.1 calibration-error points; 0.4% wrong at 95%+ confidence | decider-35b-a3b on the Decision Index calibration comparison, as reported by Marosi in 2026 | Calibration results do not establish that an individual high-confidence answer is correct |
| 18 ms median per decision | decider-2b on one B300, bf16, batch size one | Marosi’s 2026 measurement; no CUDA graphs or compilation |
| 41 ms median per decision | decider-35b-a3b on one B300, bf16, batch size one | Marosi’s 2026 measurement; no CUDA graphs or compilation |
| 132,422 requests | Decision Index edition 0.1, by multimodalart, 2026, as reported by Marosi | The article says open reproductions ran on one RTX PRO 6000 and unanswered requests counted as wrong |
| 91.5% correct on 71 held-out image questions | decider-35b text weights, as reported by Marosi in 2026 | Small, specific held-out image-question evaluation |
| 87.3% correct on 71 held-out image questions | Purpose-trained decider-2b-vision, as reported by Marosi in 2026 | Small, specific held-out image-question evaluation |
Marosi says the package’s default CUDA-graph engine is faster than the cited baseline timings, but those measurements should not be generalized to other hardware, batch sizes, or runtime settings. The Decision Index results were third-party results the author says he did not run himself. For a deployment decision, test the intended model on representative examples, measure the actual serving path, and check errors as well as confidence.
Quick Recap
A practical evaluation checklist
- Write down the allowed answers and confirm that Choice, Score, or Noul captures the decision your application actually needs.
- Test on representative cases, including ambiguous and difficult examples—not only typical inputs.
- Measure task accuracy and calibration separately; a calibrated output can still be incorrect.
- Benchmark the exact model, precision, hardware, runtime, and batch size you plan to serve.
- Define what the application should do when confidence is low or when the decision has meaningful consequences.
- Verify language support and behavior for your deployed model version, especially if inputs are not in English.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
