FuturixAIChoice
Pick the best option from a list.
Which team should handle this?
Experimental model · outputs may change between releases
Give it options. Get probabilities.
Pass in a situation and your options. Get back a calibrated probability for each one, with no free text to parse.
Illustrative examples
From a single question to a confident answer. Give it the context and your options, and it returns a calibrated probability for each one, so your software can act on the sure cases and hand the rest to a person.
- 0
- Decision Index 0.3 public score
- 0B
- Parameters
- 2–0
- Options per decision
- 0
- Context tokens
Why a decision model
Most AI decisions
are a choice.
Routing a support ticket, picking which tool an agent should call, classifying intent, scoring a response: in each case the answer is one of a known set of options.
General-purpose chat models handle this by generating text and hoping it matches one of your labels. FuturixAI Decision Model is trained for the choice itself. You pass in the options and get back a probability distribution over exactly those options.
There’s nothing to parse, and the model can’t answer with an option you didn’t give it.
A chat modelfree text
You have to parse it and hope it matches a label.
FuturixAI Decision Model● calibrated
billing0.91
support0.06
sales0.03
One probability per option you gave it. Ready to act on.
How it works
From context to
a confident decision.
Three inputs in, one probability per option out. Act on the top choice, set a confidence threshold, or send uncertain cases to a person.
- 01 / STATE
The context
A customer message, a document, a log, or the current state of an agent.
- 02 / QUESTION + OPTIONS
What to decide
The question, plus 2 to 255 candidate options. Each option can carry its own short description.
- 03 / OUTPUT
Calibrated probabilities
One probability per option. When it says 0.9, it should be right about nine times in ten.
Yes / No
A binary decision with a probability.
Was the invoice paid?
Score
Rate something on a numeric scale.
Rate this response from 0 to 5.
What it’s for
Built for the decisions
inside your software.
Anywhere your system has to choose between known options.
route()
01Routing
Send tickets, messages and requests to the right place.
intent()
02Intent classification
Work out what a user actually wants.
choose()
03Multiple choice
Pick the best answer from a fixed set.
next_tool()
04Tool selection
Help AI agents choose the right tool for the next step.
score()
05Structured scoring
Grade responses, documents or outputs on a scale.
Benchmark results Decision Index 0.3
A full public run.
Zero errors.
Every benchmark in the Decision Index 0.3 public suite, run end to end. Scores are chance-corrected on a 0–100 scale.
Public index score
0/ 100
- Scoreable requests140,178
- Answered140,148
- Unsupported30
- Errors0
- Benchmarks reported / in index43 / 37
Score by area
Chance-corrected · 0–100How it compares
The four published decision models closest to it on the public index, from the Decision Index board snapshot of 20 October 2026.
| Model | Public index | Knowledge | Language | Retrieval | Tools | Arts |
|---|---|---|---|---|---|---|
| FuturixAI Decision Model | 62.26 | 51.50 | 67.43 | 61.20 | 79.68 | 47.01 |
| Darwin-27B-ZTC | 62.63 | 44.72 | 68.46 | 60.90 | 86.69 | 53.34 |
| Sieve 31B | 62.52 | 49.91 | 65.59 | 69.55 | 78.72 | 43.50 |
| Wald-26B-A4B | 62.30 | 44.71 | 64.25 | 69.05 | 83.28 | 50.86 |
| OneJev1.6-4B | 61.95 | 45.00 | 64.40 | 68.92 | 81.50 | 49.75 |
Best score in each column
FuturixAI Decision Model has the highest Knowledge & Reasoning score in this group (51.50) and is within 0.37 points of the top public index score.
Speed
9,192 requests · single RTX PRO 6000107.5msMedian
202.5msMean
255.5msp80
Requests were sent one at a time. These latency figures come from FuturixAI’s own run; the leaderboard’s official latency check is done separately by the Decision Index maintainers.
Additional holdout
Not part of Decision Index52.56%Expected correctness
1.1454Soft-label CE · lower is better
Built for developers
The Decision SDK
is included.
The model ships with the FuturixAI Decision SDK, a typed Python interface that returns decision probabilities directly.
- Typed inputs: state, question, options and an optional description for each option
- Probabilities come back directly, with nothing to parse
- Image inputs supported
- Enforces the 8,192-token limit instead of silently cutting off long inputs
from futurix_decision import DecisionInput, load_model
model = load_model(
"/path/to/FxAI-decision-model",
family="fxai", device="cuda:0",
)
result = model.decide(DecisionInput(
state="My integration fails",
question="Which team?",
options=["billing", "support"],
candidate_descriptions=["Charges and refunds",
"Integration errors"],
))
print(result.to_dict())# try your own input
python3 example.py \
--state "I was charged twice and need a refund." \
--option billing="Charges, invoices and refunds" \
--option support="Technical integration errors" \
--option sales="Questions about purchasing"Model details
Under the hood.
- Developer
- FuturixAI
- Task
- Typed decision classification and ranking
- Architecture
- Qwen3.8-27B multimodal backbone
- Parameters
- 26.1B
- Context limit
- 8,192 tokens
- Options per decision
- 2–255
- Attention
- Noncausal full attention in full-attention layers
- Pooling
- Last token
- Output
- Calibrated option probabilities
- Languages evaluated
- Primarily English
- Training approach
- Inspired by Bosun
- License
- Apache 2.0
What you need to run it
- 01Linux, Python 3.12 and an NVIDIA CUDA GPU with bfloat16 support
- 02About 29 GB of disk for the checkpoint
- 03A 48 GB GPU is recommended for inputs up to 8,192 tokens
- 04CPU inference works, but isn’t practical for a model this size
Responsible use
Where it should
and shouldn’t be used.
Use it as one step in a workflow you have tested and a person reviews, not as the final word on decisions that matter.
01 · Scope
Not intended for
- Autonomous high-stakes decisions
- Safety-critical control systems
- Medical or legal advice
- Any use without testing on your own task and human review
02 · Evaluation
Text-only evaluation
Training and benchmark checks were text-only. Image accuracy and safety haven’t been independently evaluated yet.
03 · Cost
Cost and scoring
At 26.1B parameters, the model needs a large GPU and is costly to run. The private part of the full Decision Index score and the independent latency check are run only by the maintainers.