109 lines
3.0 KiB
Markdown
109 lines
3.0 KiB
Markdown
---
|
||
license: apache-2.0
|
||
tags:
|
||
- prisma
|
||
- coding
|
||
- cybersecurity
|
||
- reasoning
|
||
- uncensored
|
||
- agent
|
||
language:
|
||
- en
|
||
- de
|
||
- zh
|
||
library_name: transformers
|
||
pipeline_tag: text-generation
|
||
---
|
||
|
||
# Prisma-32B
|
||
|
||
**Prisma-32B** is a 32 billion parameter language model optimized for advanced coding, technical reasoning, and cybersecurity workflows. It the first Prisma Model with no security blocking. It is the second release in the **Prisma** series, following [`Prisma-0.6B`](https://huggingface.co/derprofi2431/Prisma-0.6B).
|
||
|
||
Prisma-32B is designed to be a capable, direct, and technically rigorous assistant for users who need a model that engages substantively with complex technical material.
|
||
|
||
---
|
||
|
||
## Model Details
|
||
|
||
| Property | Value |
|
||
|---|---|
|
||
| **Parameters** | 32B |
|
||
| **Architecture** | Transformer Decoder |
|
||
| **Context Length** | 32,768 tokens |
|
||
| **Languages** | English, German, Chinese (+ 20 more) |
|
||
| **License** | Apache 2.0 |
|
||
|
||
---
|
||
|
||
## Intended Use
|
||
|
||
Prisma-32B is intended for:
|
||
|
||
- **Coding assistance** — full-stack development, debugging, refactoring, code review
|
||
- **Cybersecurity research** — offensive security workflows (red team, CTF, exploit analysis) and defensive workflows (incident response, hardening, secure code review)
|
||
- **Technical writing** — documentation, system specifications, architecture
|
||
- **Research and experimentation** in controlled environments
|
||
|
||
---
|
||
|
||
## Usage
|
||
|
||
```python
|
||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
||
model = AutoModelForCausalLM.from_pretrained(
|
||
"derprofi2431/Prisma-32B",
|
||
torch_dtype="auto",
|
||
device_map="auto",
|
||
)
|
||
tokenizer = AutoTokenizer.from_pretrained("derprofi2431/Prisma-32B")
|
||
|
||
messages = [
|
||
{"role": "user", "content": "Write a port scanner in Python."}
|
||
]
|
||
inputs = tokenizer.apply_chat_template(
|
||
messages, return_tensors="pt", add_generation_prompt=True
|
||
).to(model.device)
|
||
|
||
output = model.generate(inputs, max_new_tokens=2048, temperature=0.7)
|
||
print(tokenizer.decode(output[0], skip_special_tokens=True))
|
||
```
|
||
|
||
### Recommended Sampling
|
||
|
||
| Parameter | Value |
|
||
|---|---|
|
||
| `temperature` | 0.6 – 0.8 |
|
||
| `top_p` | 0.9 |
|
||
| `top_k` | 40 |
|
||
| `repetition_penalty` | 1.05 |
|
||
|
||
---
|
||
|
||
## Quantized Versions
|
||
|
||
GGUF quantizations for local inference via Ollama and llama.cpp will be released as separate repositories.
|
||
|
||
---
|
||
|
||
## Limitations and Responsible Use
|
||
|
||
- The user is fully responsible for the content they generate and how they use it.
|
||
- The model is not aligned for general consumer-facing deployment. For production use, deploy behind an appropriate safety layer (input filtering, output classification, etc.).
|
||
- The model may reflect biases present in large-scale text corpora.
|
||
- Intended for adult, technically competent users in controlled environments.
|
||
|
||
By downloading or using this model, you agree to use it lawfully and ethically within your jurisdiction. The author assumes no liability for misuse.
|
||
|
||
---
|
||
|
||
## Citation
|
||
|
||
```bibtex
|
||
@misc{prisma32b2026,
|
||
title = {Prisma-32B},
|
||
author = {Jannik},
|
||
year = {2026},
|
||
url = {https://huggingface.co/derprofi2431/Prisma-32B}
|
||
}
|
||
``` |