README.md

---
language:
- en
license: apache-2.0
library_name: gliclass
pipeline_tag: zero-shot-image-classification
base_model: knowledgator/gliclass-v4-9b
model_name: Apex-large
tags:
- apex
- gliclass
- gliclass-v4
- zero-shot-classification
- text-classification
- multi-label-classification
- instruction-following
- reranking
- multimodal
- image-classification
---
# Apex-large

**Evidence in, decisions out.** Explore [SCX Apex](https://labs.scx.ai/apex), a family of GLiClass 4 based Large Decision Models.

**A 9B Large Decision Model for knowledge-heavy text and image decisions.**

Apex-large reads evidence and candidate answers together, using instructions, policies, or rubrics to guide its scores. Use it for more demanding classification, reasoning, and visual decision workflows.

Part of the [SCX Apex family](https://huggingface.co/collections/scx-admin/apex-family-6ad58a4ea2e0c9875ddb4061), **based on GLiClass 4**. Give the model evidence, an instruction, and the answers you would accept. It scores the candidate answers in one forward pass without generating text.

| Property | Value |
| --- | --- |
| Parameters | Approximately 9B |
| Inputs | Text and images |
| Backbone | [Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) |
| Framework | [GLiClass](https://github.com/Knowledgator/GLiClass) |
| License | Apache 2.0 |

## Use an LLM to write. Use Apex to decide.

- **Change the policy, not the model.** Supply instructions, policies, or rubrics at inference time, and change candidate answers as your workflow evolves.
- **Act on confidence.** Use calibrated label scores to automate confident decisions and escalate uncertain cases to a person or an LLM. Validate thresholds on your own data.
- **Make decisions directly.** Get structured label scores for routing, triage, moderation, ranking, extraction decisions, judging, and selecting the next step.
- **Bring images and conversations into the decision.** Score text alongside screenshots, receipts, charts, or camera frames. The GLiClass decoder-KV architecture also supports streaming classification with a per-session cache. See the [upstream streaming examples](UPSTREAM_README.md).

## Choose an Apex model

| Model | Size | Inputs | Best suited to |
| --- | --- | --- | --- |
| [Apex-flash](https://huggingface.co/scx-admin/Apex-flash) | 0.4B | Text | High-volume decisions on CPU or GPU |
| [Apex-large](https://huggingface.co/scx-admin/Apex-large) | 9B | Text and images | Knowledge-heavy and visual decisions |

Explore the [Apex family collection](https://huggingface.co/collections/scx-admin/apex-family-6ad58a4ea2e0c9875ddb4061). Both releases provide open weights under Apache 2.0.

## Quick start

Use Python 3.10+ and an up-to-date GLiClass installation:

```bash
pip install -U gliclass
```

The example uses a GPU with BF16 support. The BF16 weights alone occupy approximately 18 GB; allow additional memory for inference.

```python
import torch
from gliclass import GLiClassModel, ZeroShotClassificationPipeline

model_id = "scx-admin/Apex-large"
device = "cuda:0"
model = GLiClassModel.from_pretrained(model_id, torch_dtype=torch.bfloat16)
pipeline = ZeroShotClassificationPipeline(
    model,
    model_id,
    classification_type="multi-label",
    max_length=8192,
    device=device,
)

results = pipeline(
    "My parcel has not arrived. Please check the delivery status.",
    ["delivery issue", "refund request", "technical support"],
    threshold=0.5,
)[0]
for result in results:
    print(result["label"], result["score"])
```

Pass `prompt="your instruction or policy"` to guide the decision. Use `classification_type="single-label"` for mutually exclusive choices. Multi-label results may contain more than one label or no labels above the threshold.

### Images

Use the same pipeline with an image path, URL, or PIL image:

```python
results = pipeline(
    ["Which fruit is shown?"],
    ["strawberry", "lime", "fig", "peach"],
    images=[["fruit.jpg"]],
    classification_type="single-label",
)[0]
print(results)
```

## Limitations

Predictions depend on label wording, instructions, domain, input length, and thresholds. English is the primary language. Calibration can drift on new domains; evaluate scores and thresholds on representative data. See the [upstream model card](https://huggingface.co/knowledgator/gliclass-v4-9b) for training details, evaluations, and further limitations.

## Attribution and license

This release republishes the upstream checkpoint under the Apex name. The weights, configuration, tokenizer, and other inference assets are unchanged; no additional fine-tuning is claimed.

Based on **GLiClass 4**, developed by **Knowledgator**. Original checkpoint: [knowledgator/gliclass-v4-9b](https://huggingface.co/knowledgator/gliclass-v4-9b/tree/14d11a92b5e0bab4a166e1fce4cc0777cac27182), revision `14d11a92b5e0bab4a166e1fce4cc0777cac27182`. Distributed under the upstream **Apache License 2.0**; see [LICENSE](LICENSE). The original model card is preserved in [UPSTREAM_README.md](UPSTREAM_README.md) and at the upstream link.

Related work: [GLiClass](https://arxiv.org/abs/2508.07662) and [SCX Router](https://arxiv.org/abs/2609.02292).