3x its base model on a sealed closed-book Linux diagnostics exam (25% → 75%)
MMLU 73.7% (vs 74.0% base) — specialization without catastrophic forgetting
Built for agent swarms: an 8-agent pack runs in 15.5 GB VRAM at 1,539 tok/s aggregate on one RTX 5090 (Q8_0 GGUF via Ollama), with zero per-agent slowdown
Playful, rigorous "toon pup" persona — step-by-step reasoning with machine-checkable final answers

Why this model exists

I originally built Nova Pup as a local Linux study and troubleshooting assistant. I had collected technical books, manuals, man pages, and documentation over years of study, but I needed a more practical way to connect and apply that material while working.

The purpose was not to reproduce or redistribute those sources. It was to build a small assistant that could help explain concepts, troubleshoot systems, and apply Linux knowledge to new problems.

After testing it for my own use, I decided to release the trained model so other Linux users could experiment with it as well. The original books, PDFs, and training corpus are not included with the model.

Training

Base: InternScience Agents-A1-4B (Qwen3.5 hybrid linear-attention family)
Phase 1 — Continued pretraining: BF16 LoRA, 3,000 steps, 4,096-token packed sequences over a curated, deduplicated, secret-scrubbed Linux corpus
Phase 2 — Solver SFT: 1,200 steps on a programmatically generated, machine-verified puzzle curriculum: logic-grid deduction, bash-pipeline reconstruction, answer verification, and task decomposition — with a consistent persona layer
Adapters merged into the base weights; this repo contains the merged BF16 safetensors

Benchmarks

Measured with EleutherAI lm-evaluation-harness (0-shot loglikelihood except GSM8K 5-shot/300 samples), BF16, RTX 5090. Base model evaluated identically.

Benchmark	Base Agents-A1 4B	Nova Pup 4B
Linux closed-book exam (sealed, 16q)	25%	75%
MMLU (57 subjects)	74.0%	73.7%
ARC-Challenge (acc_norm)	54.4%	56.6%
HellaSwag (acc_norm)	71.7%	73.9%
Winogrande	69.0%	69.8%
TruthfulQA MC2	49.6%	45.9%
GSM8K (strict)	77.7%	71.0%

Known limitations (read before using)

Math regressed: GSM8K dropped 6.7 points vs base — math was not in the training mix. Pair with a calculator/code tool for arithmetic-heavy work.
TruthfulQA dipped 3.7 points; calibrate confidence accordingly.
4B parameters: built to be small, fast, and replicated in swarms — not to out-reason frontier models.
8k context in the published configuration; long-context training planned.
Distinct persona: a playful cartoon pup that verifies its work. Override the system prompt for a neutral assistant.

Usage (transformers)

python

from transformers import AutoModelForImageTextToText, AutoTokenizer
import torch

model = AutoModelForImageTextToText.from_pretrained(
    "hungrynovalabs/nova-pup-4b", dtype=torch.bfloat16, device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("hungrynovalabs/nova-pup-4b")

messages = [{"role": "user", "content": "My apt is broken after a full disk. First steps?"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
print(tokenizer.decode(model.generate(**inputs, max_new_tokens=400)[0]))

Note: the model class is Qwen3_5ForConditionalGeneration (vision-language base); use AutoModelForImageTextToText, not AutoModelForCausalLM.

Training data & licensing

Nova Pup was trained using lawfully acquired technical materials and publicly available Linux documentation. Openly licensed materials retain their respective licenses. Copyrighted materials were used only for intermediate model training and are not distributed with the model.

Base model: InternScience Agents-A1-4B (Apache-2.0). Based on Qwen3.5-4B (Apache 2.0) by the Qwen Team.