Nyx-35B

Nyx-35B

Nyx-35B is a merged full-weight instruction model with sequential LoRA fine-tuning for coding/reasoning and cybersecurity instruction following.

Model Summary

Training

Nyx-35B was trained with a two-stage sequential LoRA workflow:

1. Stage 1: CodeX pilot

   • Dataset: Modotte/CodeX-2M-Thinking
   • Rows: 20,000
   • Goal: improve coding and Python reasoning behavior

2. Stage 2: Cyber specialization

   • Dataset: jmtss/cyber-security-instruct-3k
   • Rows: 3,678
   • Effective batch size: 32
   • Steps: 115
   • Learning rate: 5e-5
   • Final train loss: 1.511

The final uploaded model is a merged model:

base model + Stage 1 CodeX adapter + Stage 2 Cyber adapter

Recommended Hardware

For full-precision inference, the model needs more than the raw 70 GB weight size because serving also requires runtime memory and KV cache.

Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "jmtss/Nyx-35B"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

messages = [
    {"role": "user", "content": "Write a short Python function that checks if a URL uses HTTPS."}
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    temperature=0.2,
    do_sample=True,
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

vLLM

For production serving, use vLLM if your environment supports this Qwen3.6 MoE architecture:

vllm serve jmtss/Nyx-35B \
  --trust-remote-code \
  --dtype bfloat16 \
  --max-model-len 4096

Increase --max-model-len only if your GPU has enough free memory for KV cache.

Intended Use

Nyx-35B is intended for:

• Python and software engineering assistance
• Defensive cybersecurity education
• MITRE ATT&CK-style concept explanation
• Security documentation and analysis support
• General technical instruction following

Safety and Limitations

• This model has not been formally benchmarked beyond training loss and basic sanity prompts.
• Cybersecurity outputs should be reviewed by a qualified human before operational use.
• The model may produce incorrect, outdated, or incomplete security guidance.
• The cybersecurity tuning is intended for defensive, educational, and authorized research contexts.
• Do not use this model for unauthorized access, credential theft, malware deployment, evasion, or other harmful activity.
• The model may include thinking-style prefaces in responses because of the base and training data style.

Training Artifacts

The uploaded repository contains the merged full model only. Intermediate LoRA adapters and training checkpoints were not included in this repository.

License

This model is a derivative of the listed base model and datasets. Use is subject to the terms of the base model, datasets, and any applicable licenses. Verify compatibility for your use case before commercial or production deployment.