> For the complete documentation index, see [llms.txt](https://docs.nvcl.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.nvcl.ai/concepts/quality.md).

# Why quality holds

Smaller weights only help if answer quality holds. We check that two ways: a quick closeness check after every change, and full task benchmarks against the original model.

The result on Qwen3.8-27B: the default tier, Standard (about 7.13 bits per weight), shows no measurable quality loss against the full-precision model, through the chat template, on our 1,015-task suite (the differences stay within its run-to-run noise), on standard benchmarks, on multi-turn instructions in eight languages and on long context up to 128K tokens. Every score, with its conditions: [Quality](/benchmarks/quality.md).

## The quick check: how far predictions drift

For every next token, a model gives a probability to each token it knows: "the" 41%, "a" 23%, and so on. We feed the original 16-bit model and a compressed one the same text and compare those probabilities.

The result is one number, the **KL divergence**. 0 means identical; larger means the compressed model drifts further from the original. Alongside it we count how often both models put the same token first.

Every weight format and every change to NVCL AI's GPU code goes through this check.

**The noise floor.** Even Exact, the original weights bit for bit, does not score exactly 0. The GPU adds numbers in a slightly different order than the reference does, and that alone moves the result a little. So a score near Exact's means "as close as this check can see".

{% hint style="info" %}
KL is more sensitive than a task score: it sees a change in the model's probabilities even when every answer stays the same. That makes it a fast alarm after every change. Whether answers got worse is decided by the task benchmarks below.
{% endhint %}

## The task check: same questions, compared one by one

The task-level check runs the same questions on NVCL AI and on the original model served by vLLM or SGLang, and compares them question by question with paired statistical tests. How the results are read: [How quality is compared](/concepts/benchmarking.md#how-quality-is-compared).

## Formats and their quality class

A *quality class* groups setups that stay about equally close to the original model.

A format's bit-width alone says nothing about its quality. Quality comes from the checks above. Each tier's size, quality class and evidence are in [Weight formats](/reference/formats.md).