---
language: en
tags:
- recursive-seed-ai
- grok
- distillation
- fine-tuned
- custom-architecture
pipeline_tag: text-generation
library_name: transformers
datasets:
- WithinUsAI/Grok4.4_heavy_max_distill_god_seed_25k
- 11-47/god_agent_grok4.4_cot_traces_20k
- WithinUsAI/grok_frontier_dataset_v3_100k
- WithinUsAI/Grok_4.4_Distilled
---

# GOD_Agent_Grok4.4 — Recursive Seed AI

Fine-tuned distillation of the Recursive Seed AI architecture on Grok 4.4 frontier reasoning data.

## Model Details

- **Architecture**: Recursive Seed AI (custom, see modeling_recursive_seed.py)
- **Parameters**: 158M
- **Context Window**: 60,000
- **Modalities**: Text (+ vision/audio encoder stubs)
- **License**: Gated / Manual Approval

## Training

| Parameter | Value |
|-----------|-------|
| Datasets | grok_frontier_dataset_v3_100k, Grok_4.4_Distilled, Grok4.4_heavy_max_distill_god_seed_25k |
| Total Examples | 141,314 |
| Training Steps | 2,000 |
| Batch Size | 1 (grad accum 8) |
| Learning Rate | 1e-4 with 100-step warmup |
| Optimizer | AdamW (weight_decay=0.01) |
| Final Loss | 0.52 |
| Duration | ~1 hour (CPU) |

### Loss Curve

| Step | Loss |
|------|------|
| 100 | 1.32 |
| 500 | 1.01 |
| 1000 | 0.74 |
| 1500 | 0.63 |
| 2000 | 0.52 |

## Usage

```python
from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer

config = AutoConfig.from_pretrained("11-47/GOD_Agent_Grok4.4", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    "11-47/GOD_Agent_Grok4.4",
    config=config,
    trust_remote_code=True,
    torch_dtype="auto",
)
tokenizer = AutoTokenizer.from_pretrained("11-47/GOD_Agent_Grok4.4", trust_remote_code=True)

inputs = tokenizer("User: What is recursive seed AI?
Assistant: ", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## Gated Access

This model requires manual approval to access.