---
library_name: transformers
tags: [hrm, moe, hierarchical-reasoning, custom-architecture]
---

# Hierarchical Reasoning Model (HRM)

Custom MoE language model with 3-level hierarchical reasoning and DeepSeek-V3 memory strategies.

**Architecture:** 3 levels · 16 experts (4 active) · MLA attention · Hierarchical memory  
**Parameters:** ~350M total, ~45M active per token  
**Trained on:** OpenHermes-2.5

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Scorched2/shader-v2", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("Scorched2/shader-v2")

inputs = tokenizer("### Instruction:\nExplain AI.\n\n### Response:\n", return_tensors="pt")
out = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(out[0]))
```