Release FxAI decision model

f346832 6 days ago

6.27 kB

---
tags:
  - decision-model
  - classification
  - text-classification
license: apache-2.0
language:
  - en
library_name: transformers
pipeline_tag: text-classification
inference: false
---
# FxAI Decision Model

![FxAI Decision Model](assets/fxai-decision-model-banner.png)

FxAI Decision Model is a trained decision classifier on the
[Qwen/Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B) architecture.
Given a state, a question and up to 255 candidate options, it returns a
calibrated probability distribution over those options.

The repository contains the complete 26.1B checkpoint and the Futurix
Decision SDK.

The model completed the **Decision Index 0.3 public suite** with a public
score of **62.26**.

## Model details

- **Task:** typed decision classification and ranking
- **Architecture:** Qwen3.8-27B multimodal backbone
- **Parameters:** 26.1B
- **Context limit:** 8,192 tokens through the bundled decision interface
- **Options:** 2–255 per question through the Decision SDK
- **Attention:** noncausal full attention in full-attention layers
- **Pooling:** last token
- **Output:** calibrated option probabilities
- **Languages evaluated:** primarily English

## Intended use

This model is intended for routing, intent classification, multiple-choice
decisions, tool selection and structured scoring.

It is not intended for autonomous high-stakes decisions, safety-critical
control, medical or legal advice, or use without application-specific
evaluation and human review.

## Inference

This repository includes the Futurix Decision SDK under `futurix_decision/`.
Use that runtime for decision probabilities. `transformers.AutoModel` alone
does not return them. The Hugging Face Inference widget is disabled.

### Requirements

- Linux, Python 3.12 and an NVIDIA CUDA GPU with bfloat16 support.
- The checkpoint is about 29 GB. A 48 GB GPU is recommended for inputs up
  to 8,192 tokens.
- CPU inference is supported but is not a practical way to run this 27B
  model.

Clone the repository with Git LFS, create an isolated environment and
install the runtime used for verification:

```bash
git lfs install
git clone https://huggingface.co/Futurix-AI/FxAI-decision-model
cd FxAI-decision-model
python3 -m venv .venv
source .venv/bin/activate
python3 -m pip install --upgrade pip
python3 -m pip install -r requirements.txt
python3 example.py --device cuda:0
```

To try another input:

```bash
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"
```

### Python API

```python
from pathlib import Path
from futurix_decision import DecisionInput, load_model
checkpoint = Path("/path/to/FxAI-decision-model")
model = load_model(
    checkpoint, family="fxai", device="cuda:0",
    max_image_pixels=512 * 512,
)
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())
```

The SDK supports typed `choice`, `noul`, and `score` decisions with 2–255
options. The FxAI runtime supports image inputs, enforces an 8,192-token
limit without silent truncation, and returns calibrated option
probabilities.

Convenience methods use the same runtime:

```python
choice = model.choice(
    state="The customer was charged twice.",
    question="Which team should handle this?",
    choices=["billing", "support"],
)
binary = model.noul(state="The invoice is paid.", question="Was it paid?")
score = model.score(
    state="The response is mostly correct.",
    question="Rate its quality.",
    min_score=0,
    max_score=5,
)
```

Training and benchmark checks were **text-only**. Image accuracy and safety
have not been independently evaluated.

## Training details

The training strategy is inspired by Bosun. The model is a trained decision
classifier on the Qwen3.8-27B architecture.

## Decision Index 0.3 public benchmark

The official scorer covered **140,178** scoreable requests over 43 reported
benchmarks: 140,148 answered, 30 unsupported, zero errors and zero pending.
The public index counts 37 benchmarks.

| Result | FxAI-decision-model |
|---|---:|
| **Public Decision Index** | **62.26** |
| Knowledge & Reasoning | 51.50% |
| Language Understanding | 67.43% |
| Retrieval & Classification | 61.20% |
| Tools & Automation | 79.68% |
| Arts & Human Taste | 47.01% |

### Published comparison

The four published decision models nearest this public index, from the
Decision Index board snapshot of 20 October 2026, 14:13 UTC
([Decision Index board](https://huggingface.co/spaces/multimodalart/jev-decision-index)).
Area columns are the same chance-corrected public scores, on a 0–100 scale.

| Model | Public index | Knowledge | Language | Retrieval | Tools | Arts |
|---|---:|---:|---:|---:|---:|---:|
| **FxAI 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 |

Candidate-only machine-readable results are available in `BENCHMARKS.json`.

Evaluation ran in bounded sequential and throughput-batched chunks on one
RTX PRO 6000. The mixed full-run timing fields are throughput measurements,
**not** a valid single-request leaderboard latency claim.
On the separate 9,192-request sequential component run, median / mean / p80
were **107.5 / 202.5 / 255.5 ms**. The maintainers must independently check
the private latency gate.

### Additional independent holdout

On a separate holdout, expected correctness was **52.56%** and soft-label
cross-entropy was **1.1454** (lower is better). This result is not part of
Decision Index.

## Limitations and release status

The 27B checkpoint and GPU requirement are costly. The published figure is
the public Decision Index score. The private portion of the Full score and
the independent latency gate are available only to the maintainers.

	---
	tags:
	- decision-model
	- classification
	- text-classification
	license: apache-2.0
	language:
	- en
	library_name: transformers
	pipeline_tag: text-classification
	inference: false
	---

	# FxAI Decision Model

	![FxAI Decision Model](assets/fxai-decision-model-banner.png)

	FxAI Decision Model is a trained decision classifier on the
	[Qwen/Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B) architecture.
	Given a state, a question and up to 255 candidate options, it returns a
	calibrated probability distribution over those options.

	The repository contains the complete 26.1B checkpoint and the Futurix
	Decision SDK.

	The model completed the Decision Index 0.3 public suite with a public
	score of 62.26.

	## Model details

	- Task: typed decision classification and ranking
	- Architecture: Qwen3.8-27B multimodal backbone
	- Parameters: 26.1B
	- Context limit: 8,192 tokens through the bundled decision interface
	- Options: 2–255 per question through the Decision SDK
	- Attention: noncausal full attention in full-attention layers
	- Pooling: last token
	- Output: calibrated option probabilities
	- Languages evaluated: primarily English

	## Intended use

	This model is intended for routing, intent classification, multiple-choice
	decisions, tool selection and structured scoring.

	It is not intended for autonomous high-stakes decisions, safety-critical
	control, medical or legal advice, or use without application-specific
	evaluation and human review.

	## Inference

	This repository includes the Futurix Decision SDK under `futurix_decision/`.
	Use that runtime for decision probabilities. `transformers.AutoModel` alone
	does not return them. The Hugging Face Inference widget is disabled.

	### Requirements

	- Linux, Python 3.12 and an NVIDIA CUDA GPU with bfloat16 support.
	- The checkpoint is about 29 GB. A 48 GB GPU is recommended for inputs up
	to 8,192 tokens.
	- CPU inference is supported but is not a practical way to run this 27B
	model.

	Clone the repository with Git LFS, create an isolated environment and
	install the runtime used for verification:

	```bash
	git lfs install
	git clone https://huggingface.co/Futurix-AI/FxAI-decision-model
	cd FxAI-decision-model
	python3 -m venv .venv
	source .venv/bin/activate
	python3 -m pip install --upgrade pip
	python3 -m pip install -r requirements.txt
	python3 example.py --device cuda:0
	```

	To try another input:

	```bash
	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"
	```

	### Python API

	```python
	from pathlib import Path

	from futurix_decision import DecisionInput, load_model

	checkpoint = Path("/path/to/FxAI-decision-model")
	model = load_model(
	checkpoint, family="fxai", device="cuda:0",
	max_image_pixels=512 * 512,
	)
	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())
	```

	The SDK supports typed `choice`, `noul`, and `score` decisions with 2–255
	options. The FxAI runtime supports image inputs, enforces an 8,192-token
	limit without silent truncation, and returns calibrated option
	probabilities.

	Convenience methods use the same runtime:

	```python
	choice = model.choice(
	state="The customer was charged twice.",
	question="Which team should handle this?",
	choices=["billing", "support"],
	)
	binary = model.noul(state="The invoice is paid.", question="Was it paid?")
	score = model.score(
	state="The response is mostly correct.",
	question="Rate its quality.",
	min_score=0,
	max_score=5,
	)
	```

	Training and benchmark checks were text-only. Image accuracy and safety
	have not been independently evaluated.

	## Training details

	The training strategy is inspired by Bosun. The model is a trained decision
	classifier on the Qwen3.8-27B architecture.

	## Decision Index 0.3 public benchmark

	The official scorer covered 140,178 scoreable requests over 43 reported
	benchmarks: 140,148 answered, 30 unsupported, zero errors and zero pending.
	The public index counts 37 benchmarks.

	\| Result \| FxAI-decision-model \|
	\|---\|---:\|
	\| Public Decision Index \| 62.26 \|
	\| Knowledge & Reasoning \| 51.50% \|
	\| Language Understanding \| 67.43% \|
	\| Retrieval & Classification \| 61.20% \|
	\| Tools & Automation \| 79.68% \|
	\| Arts & Human Taste \| 47.01% \|

	### Published comparison

	The four published decision models nearest this public index, from the
	Decision Index board snapshot of 20 October 2026, 14:13 UTC
	([Decision Index board](https://huggingface.co/spaces/multimodalart/jev-decision-index)).
	Area columns are the same chance-corrected public scores, on a 0–100 scale.

	\| Model \| Public index \| Knowledge \| Language \| Retrieval \| Tools \| Arts \|
	\|---\|---:\|---:\|---:\|---:\|---:\|---:\|
	\| FxAI 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 \|

	Candidate-only machine-readable results are available in `BENCHMARKS.json`.

	Evaluation ran in bounded sequential and throughput-batched chunks on one
	RTX PRO 6000. The mixed full-run timing fields are throughput measurements,
	not a valid single-request leaderboard latency claim.
	On the separate 9,192-request sequential component run, median / mean / p80
	were 107.5 / 202.5 / 255.5 ms. The maintainers must independently check
	the private latency gate.

	### Additional independent holdout

	On a separate holdout, expected correctness was 52.56% and soft-label
	cross-entropy was 1.1454 (lower is better). This result is not part of
	Decision Index.

	## Limitations and release status

	The 27B checkpoint and GPU requirement are costly. The published figure is
	the public Decision Index score. The private portion of the Full score and
	the independent latency gate are available only to the maintainers.