my-ai-stack/Stack-4.0-Qwen-3B-Merged

Stack 4.0 Omni-Nexus

Merged · 3B Parameters · Sovereign Agentic Infrastructure

---

Stack 4.0 Omni-Nexus — Merged

Model ID: my-ai-stack/Stack-4.0-Qwen-3B-Merged

A 3-billion parameter instruction-tuned coding model, fully merged from Qwen2.5-Coder-3B-Instruct with 55,000 agentic tool-use conversations baked in. This is the standalone version — no adapter needed, runs directly on any compatible hardware.

Performance Benchmarks

Benchmark	Score	Notes
HellaSwag (acc_norm)	74.0%	50-sample eval
ARC-Challenge (acc_norm)	52.0%	50-sample eval
Internal coding sample	10/10	All valid Python produced

Key Metrics

Metric	Value
Parameters	3B
Training loss (final)	0.1411
Training steps	1,000
Hardware	GCP Tesla V100 16GB
Training time	~10 hours

Why Merged?

The merged version ships the full model in a single file — no LoRA adapters, no base model dependency. Deploy anywhere that supports Hugging Face Transformers.

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

MODEL = "my-ai-stack/Stack-4.0-Qwen-3B-Merged"

tokenizer = AutoTokenizer.from_pretrained(MODEL, trust_remote_code=True)
tokenizer.pad_token = tokenizer.eos_token

model = AutoModelForCausalLM.from_pretrained(
    MODEL, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True
)
model.eval()

messages = [{"role": "user", "content": "Write a quicksort in Python"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=512, temperature=0.7)

print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Training Details

Parameter	Value
Method	QLoRA → Merged
LoRA rank	16
Trainable params	7.3M / 3.1B (0.24%)
Batch size	1
Grad accumulation	16
Max length	512
Learning rate	2e-4
Optimizer	AdamW (bf16)
Hardware	GCP V100 16GB

Limitations

• 3B model — smaller than 7B models; less capable on complex multi-step reasoning
• English-optimized — other language performance may vary
• Tool execution — tool calls are generated but actual execution requires an agent loop in your application

See Also

• LoRA Adapter version — smaller, needs base model
• Training dataset
• Stack 3.0 (7B)

Citation

@misc{stack-4-merged-2026,
  title={Stack 4.0 Omni-Nexus — Merged},
  author={Stack AI Team},
  year={2026},
  url={https://huggingface.co/my-ai-stack/Stack-4.0-Qwen-3B-Merged}
}