---
base_model: qvac/VisionPsy-Nano-460M
base_model_relation: quantized
license: apache-2.0
language:
  - en
tags:
  - vision-language-model
  - gguf
  - quantization
  - llama.cpp
---

# VisionPsy-Nano TriStack

**A flagship challenger at a smaller package: 9W/7L/1T over 17 displayed rows vs QVAC q4_k_m-imat (same-harness re-run) at −5.4 MiB (−1.4%) — the cleanest measured balance of size, quality and speed in the release, all three techniques stacked.**

A GGUF quantized derivative of QVAC's [VisionPsy-Nano-460M](https://huggingface.co/qvac/VisionPsy-Nano-460M) built with **P1 (domcal) × P4 (embguard) × P10 (projcomp)** — all three surviving quantization-side techniques stacked in one package. This is not a new trained checkpoint: the ~460M-parameter architecture is unchanged. Part of the VisionPsy-Nano release collection — see the Links section below.

## Exact runtime pair

| Component | File | Bytes |
|---|---|---:|
| LM | `visionpsy-nano-460m-q4_k_m-domcal-embguard.gguf` | 320862016 |
| mmproj | `MY_q4_0-domcal-mmproj-q6.gguf` | 85355072 |

Package size: **387.4 MiB**. Use only this LM/mmproj mapping. SHA-256 checksums for both files ship in this repository as [SHA256SUMS](https://huggingface.co/simoneschiavoi/VisionPsy-Nano-TriStack/blob/main/SHA256SUMS) — verify after download with `sha256sum -c SHA256SUMS`.

## Model at a glance

| | |
|---|---|
| Base model | QVAC VisionPsy-Nano-460M (~460M parameters; SigLIP2 vision encoder + SmolLM2-360M backbone) |
| Techniques | **P1 (domcal)** — domain-calibrated imatrix quantization; **P4 (embguard)** — input embedding tensor (`token_embd`) held at Q8_0; **P10 (projcomp)** — multimodal projector compression q8 → q6_K |
| Quantization | q4_k_m LM (byte-identical to the DomCal-EmbGuard artifact) + q6_K mmproj (byte-identical to the DomCal-Slim one) — a composition of two already-verified artifacts, no new quantization step |
| Calibration data | VQAv2-train + TextVQA-train, ChatML-formatted; training splits only — never benchmark test data, no evaluation images |
| Total size | **387.4 MiB** = 306.0 MiB LM + 81.4 MiB mmproj — −5.4 MiB (−1.4%) vs QVAC q4_k_m-imat (392.8) |
| Headline | **9W/7L/1T over 17 displayed rows vs QVAC q4_k_m-imat (same-harness re-run)**; normalized **61.64** vs 61.54 at a smaller package; best measured POPE (87.89) and ScienceQA (85.13) across all six ledger models; **13W/4L/0T vs QVAC q4_0 (same-harness)** |
| Observed speed | 795.4 ms/item — 2nd fastest observed in this 200-item solo sweep, faster than every QVAC configuration (q4_0 803.5, imat 813.2) |

## Why this build exists

The study's conclusion: the same trained weights become meaningfully different deployment tools once calibration and precision are assigned deliberately. TriStack is the balanced assignment — the only package that stacks all three surviving techniques. The projector sits at its measured q6_K floor (−21.5% projector, quality-neutral at pooled p = 0.49), the LM carries the DomCal-EmbGuard configuration, and the composition was verified as its own artifact: against its EmbGuard parent (identical LM, q8 projector) it measured parity — mean delta ≈ −0.36 point over 16 common cells — at −5.4% package. The result is a build that beats the flagship's score at a smaller size and a lower observed latency, with the best ScienceQA of the study (85.13).

## Vs the two rulers

The primary ruler is QVAC q4_k_m-imat, QVAC's flagship build — re-run in this same harness with a hash-pinned full-17 record (judged rows scored with a qwen3.6-27b API judge, a reconstruction of QVAC's judging protocol validated within ±1 pt of their published card on 6/8 judged benchmarks; QVAC's own card numbers remain labeled context, never medaled). QVAC q4_0 is the secondary same-harness ruler. Negative deltas = smaller/faster.

| | QVAC q4_k_m-imat (same-harness re-run) | QVAC q4_0 (same-harness) |
|---|---:|---:|
| W/L/T over 17 displayed rows | **9W/7L/1T** | **13W/4L/0T** |
| Package size Δ (387.4 MiB) | −5.4 MiB (−1.4%) | +39.7 MiB (+11.4%) |
| Speed Δ (795.4 ms/item) | −17.8 ms/item (−2.2%) | −8.1 ms/item (−1.0%) |

## Measured results

Wins land on reasoning and knowledge: ScienceQA **85.13** (highest measured), MathVista **47.8 vs 47.1**, MMVet **33.5 vs 32.8**, MMMU **31.3 vs 29.3**, plus MM-IFEval@2048 **41.5 vs 41.1** under the identical 2048-token protocol. The named losses stay visible: QVAC q4_k_m-imat keeps OCRBench, DocVQA, ChartQA and InfoVQA; QVAC q4_0 keeps RealWorldQA, MME, MMMU dev and MMVet. The nine wins against the flagship come at −1.4% package and −2.2% observed latency — the balanced point of the release.

## Limitations

- **Exploratory scope:** one seed (17), one harness, one GPU. Counts are descriptive rows, not universal-superiority claims.
- **Composition note:** the LM and projector are byte-identical to two sibling artifacts; the package was still benchmarked as its own hash-pinned artifact across the full 17-benchmark suite.
- **Judged rows:** qwen3.6-27b via OpenRouter — an attempted same-model reconstruction of QVAC's judge, not their exact serving protocol.
- Parity with the DomCal-EmbGuard parent is a measured mean (−0.36 pt); the largest single-cell delta is MMVet −2.11. DomCal-EmbGuard holds the maximum-quality end.

## Links

This model is one of four verified VisionPsy-Nano GGUF packages released together under the [simoneschiavoi Hugging Face namespace](https://huggingface.co/simoneschiavoi).

- Project website (full interactive research write-up): <https://simoneschiavoi.github.io/visionpsy-optimization/>
- Benchmarks (full same-harness ledger, all models × 17 benchmarks): <https://simoneschiavoi.github.io/visionpsy-optimization/#benchmarks>
- Hugging Face namespace (all four packages): <https://huggingface.co/simoneschiavoi>
- Sibling models:
  - [VisionPsy-Nano DomCal](https://huggingface.co/simoneschiavoi/VisionPsy-Nano-DomCal)
  - [VisionPsy-Nano DomCal-Slim](https://huggingface.co/simoneschiavoi/VisionPsy-Nano-DomCal-Slim)
  - [VisionPsy-Nano DomCal-EmbGuard](https://huggingface.co/simoneschiavoi/VisionPsy-Nano-DomCal-EmbGuard)

## License and attribution

Apache-2.0 derivative. The Apache-2.0 NOTICE distributed with the artifact must be retained, and QVAC's VisionPsy-Nano-460M must be attributed as the base model. Build evidence, evaluation ledger, and reproduction scripts: <https://github.com/simoneschiavoi/visionpsy-optimization>. All four release packages: <https://huggingface.co/simoneschiavoi>.