README.md

---
license: apache-2.0
base_model: qvac/VisionPsy-Nano-460M
pipeline_tag: image-text-to-text
library_name: transformers
tags:
  - vision-language-model
  - nanovlm
  - chart-understanding
  - ocr
  - crypto
  - launchpad
  - stable-mainnet
  - fefer
  - pegd-fun
language:
  - en
---

<p align="center">
  <img src="fefer-512.png" width="180" alt="FEFER AI" />
</p>

# FEFER-AI-460M 🦖

**FEFER** is the resident chart analyst of [pegd.fun](https://pegd.fun) — the stablecoin
launchpad on **Stable mainnet (chain id 988)**. This model is a FEFER-tuned build of
[`qvac/VisionPsy-Nano-460M`](https://huggingface.co/qvac/VisionPsy-Nano-460M)
(Tether AI Research), a ~460M-parameter vision-language model (SigLIP2 vision encoder +
SmolLM2-360M LM, nanoVLM architecture) that punches far above its size on chart
understanding, OCR, and visual instruction following.

FEFER reads **one thing and one thing only**: terminal-style candlestick charts rendered by
the pegd.fun indexer for Stable-988 launchpad tokens. Each chart has verified indexer facts
(market cap, 24h volume, graduation progress) printed directly on the image, so the model
cross-checks pixels against ground truth instead of inventing numbers.

## What it's for

- Reading pegd.fun 1024×512 candle charts: trend, momentum, volume behavior, drawdowns
- Answering holder questions in the FEFER persona (plain language, risk note always included)
- Explaining launchpad mechanics it was trained on: fixed 1B supply, permanently locked
  Uniswap V3 1% USDT0 pools, 3,000 USDT0 opening FDV, 9,000 USDT0 graduation, 80/20
  creator/platform fee split

## What it's NOT for

- Price prediction or financial advice — FEFER always says so, by training and by prompt
- Multi-image reasoning (single-image by architecture)
- Anything outside Stable mainnet 988 — other chains are out of scope on purpose

## Quickstart (transformers)

```python
from transformers import AutoModelForImageTextToText, AutoProcessor
from PIL import Image

repo = "feferai/FEFER-AI-460M"  # replace after publishing
model = AutoModelForImageTextToText.from_pretrained(repo, trust_remote_code=True)
processor = AutoProcessor.from_pretrained(repo, trust_remote_code=True)

chart = Image.open("fefer_chart.png")  # a pegd.fun /api/fefer/chart/<token>.png render
messages = [{"role": "user", "content": [
    {"type": "image", "image": chart},
    {"type": "text", "text": "Read this chart. Trend, volume, and graduation status?"},
]}]
inputs = processor.apply_chat_template(messages, add_generation_prompt=True,
                                       tokenize=True, return_dict=True, return_tensors="pt")
out = model.generate(**inputs, max_new_tokens=320)
print(processor.batch_decode(out[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True)[0])
```

## Serving for pegd.fun

Any OpenAI-compatible endpoint works. Then point the launchpad at it:

```bash
# GPU
vllm serve feferai/FEFER-AI-460M --trust-remote-code --port 8000
# or CPU (no GPU needed at 460M)
python fefer-ai/serve_cpu.py --model feferai/FEFER-AI-460M --port 8008

# pegd.fun server env
FEFER_AI_URL=http://127.0.0.1:8000/v1
FEFER_AI_MODEL=feferai/FEFER-AI-460M
```

## Training

LoRA fine-tune on the language-model attention projections (vision tower frozen), on
chart/answer pairs harvested from live pegd.fun launches by `fefer-ai/build_dataset.mjs`
— every sample is a real indexer-rendered chart paired with facts computed from on-chain
trade history, so supervision is grounded, not synthetic guesswork. Merged with
`merge_and_unload` for standalone serving.

## Limitations & bias

460M parameters is small: FEFER is a sharp chart *reader*, not an oracle. It can misread
dense wicks, and it inherits any biases of the base model. Launchpad tokens are volatile;
nothing this model outputs is financial advice, and it is trained to say exactly that.

## License & attribution

Apache 2.0, same as the base model. Built on **VisionPsy-Nano-460M by Tether AI Research
(qvac)** — all credit for the base architecture and pretraining to them. FEFER branding,
chart pipeline, dataset tooling, and fine-tune by the pegd.fun team.