---

base_model: NAME0x0/AVA-v2

license: apache-2.0

language:

- en

pipeline_tag: text-generation

tags:

- gguf

- llama.cpp

- ollama

- lm-studio

- qlora

- low-resource

- reasoning

quantized_by: NAME0x0

---

<p align="center">

<img src="https://raw.githubusercontent.com/NAME0x0/AVA/main/AVA_logo.png" alt="AVA logo" width="160" />

</p>

AVA v2 — GGUF

Ready-to-run GGUF builds of AVA v2, a

2B reasoning model fine-tuned entirely on a single 4 GB laptop GPU. **82.0%

ARC-Challenge, 92.0% ARC-Easy, 59.2% MMLU** on a 17-benchmark / 16,872-task

full evaluation (report).

Works with llama.cpp, Ollama, LM Studio, Jan, KoboldCpp — no Python, no GPU

required.

Files

All sub-8-bit quants are built with an importance matrix calibrated on the

model's own training distribution (reasoning, math, science, instruction

following) — the same idea behind Google's Gemma QAT releases: keep the small

quants as close to reference quality as possible.

Measured quality cost vs the Q8_0 reference (perplexity on a held-out slice

of the training distribution, 512-token context — lower is better):

| File | Size | RAM needed | PPL | vs Q8_0 | Use when |

|---|---|---|---|---|---|

| AVA-v2-IQ4_XS.gguf | 1.11 GB | ~1.6 GB | 2.5347 | +2.0% | Tightest fit — old laptops, SBCs |

| AVA-v2-Q4_0.gguf | 1.12 GB | ~1.6 GB | 2.5244 | +1.6% | ARM/AVX-optimized CPU inference |

| AVA-v2-Q4_K_M.gguf | 1.19 GB | ~1.7 GB | 2.4907 | +0.25% | Recommended default |

| AVA-v2-Q5_K_M.gguf | 1.31 GB | ~1.8 GB | — | — | Better quality, still small |

| AVA-v2-Q8_0.gguf | 1.87 GB | ~2.4 GB | 2.4844 | reference | Matches the published eval |

Quick start

Ollama

ollama run hf.co/NAME0x0/AVA-v2-GGUF:Q4_K_M

llama.cpp

llama-cli -m AVA-v2-Q4_K_M.gguf -ngl 99 --temp 0.7 \
  -p "Explain why ice floats on water."

LM Studio / Jan

Search for NAME0x0/AVA-v2-GGUF in the model browser and download a file.

Chat format

Qwen3.5 ChatML-style template (embedded in the GGUF — runtimes apply it

automatically):

<|im_start|>user
{your prompt}<|im_end|>
<|im_start|>assistant

Benchmarks (Q8_0, full sets, 95% Wilson CI)

| Benchmark | n | Accuracy |

|---|---|---|

| ARC-Easy | 2,376 | 92.0% |

| ARC-Challenge | 1,172 | 82.0% |

| PIQA | 1,838 | 75.9% |

| BoolQ | 3,270 | 75.0% |

| MMLU (5-shot) | 14,042 | 59.2% |

| GSM8K (greedy / k=5) | 1,319 / 200 | 35.3% / 44.0% |

Full 17-benchmark table and protocol: RESULTS_REPORT_V2_FULL.md.

At 2B parameters, AVA v2's ARC-Challenge (82.0%) sits ahead of Llama 3.2

3B-Instruct (78.6%) and within two points of Phi-4-mini 3.8B (83.7%) — models

trained with cluster-scale compute. AVA v2 was trained in 100 minutes on one

4 GB laptop GPU.

Provenance

• Adapter + training details: NAME0x0/AVA-v2
• Base model: Qwen/Qwen3.5-2B (Apache 2.0)
• Everything reproducible: github.com/NAME0x0/AVA — corpus builders, training configs, eval harness, and this quantization pipeline are all in the repo.

Citation

@misc{ava-v2-2026,
  title={AVA v2: QLoRA Fine-tuning Under Extreme VRAM Constraints},
  author={Muhammad Afsah Mumtaz},
  year={2026},
  url={https://github.com/NAME0x0/AVA}
}