> 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/reference/formats.md).

# Weight formats

NVCL AI stores model weights in its own quantization formats, offered as four tiers, and Small for Macs with Apple Silicon. The default, Standard, has fewer bits than FP8 (the official 8-bit checkpoint) and the quality of BF16 (the original full-precision model). Need the original bit for bit? Switch to Exact.

{% hint style="success" %}
**Lossless quality.** Standard shows no measurable quality loss on our 1,015-task suite: 732 / 732 / 730 over three runs against 726 for BF16 on vLLM 0.31 and SGLang 0.5.21, differences that stay within its run-to-run noise (Qwen3.8-27B on H200; [Quality](/benchmarks/quality.md)). Exact serves the original weights bit for bit.
{% endhint %}

<figure><picture><source srcset="https://3133985975-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FDZCnCH3aMU9wYLt2Z4wR%2Fuploads%2Fgit-blob-7a581e182df714ca4fa03aafb70b8c872a5ab268%2Fchart-bits-per-weight-dark.svg?alt=media" media="(prefers-color-scheme: dark)"><img src="https://3133985975-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FDZCnCH3aMU9wYLt2Z4wR%2Fuploads%2Fgit-blob-73545be405714f3355c3eb5c2b3f57c428d8cf16%2Fchart-bits-per-weight-light.svg?alt=media" alt="Bar chart of bits per weight: BF16, the original model, 16; Exact 12; FP8, the official 8-bit model, 8; Standard, the default, about 7.13; Throughput about 7; Compact about 5.8. BF16, Exact and Standard are in the full-precision quality class; FP8, Throughput and Compact in the FP8 quality class."></picture><figcaption><p>Bits per weight of the four tiers of Qwen3.8-27B, with BF16 and FP8 for reference; the colour is the quality class. Quality of the Throughput tier: FP8 class, by our full 1,015-task run.</p></figcaption></figure>

| Tier                   | Bits per weight | Size vs BF16       | Quality class                     | Evidence                                                                                  |
| ---------------------- | --------------: | ------------------ | --------------------------------- | ----------------------------------------------------------------------------------------- |
| **Standard** (default) |      about 7.13 | about 2.2x smaller | BF16 (full precision)             | 1,015-task suite, standard benchmarks, long documents: level with BF16                    |
| **Compact**            |       about 5.8 | about 2.8x smaller | FP8                               | 1,015-task suite: no significant difference from the official FP8 checkpoint on vLLM 0.32 |
| **Throughput**         |         about 7 | same as Standard   | FP8                               | 1,015-task suite: no significant difference from the official FP8 checkpoint on vLLM 0.32 |
| **Exact**              |              12 | a quarter smaller  | The original weights, bit for bit | Every weight decodes to its original value                                                |

What a quality class is: [Why quality holds](/concepts/quality.md#formats-and-their-quality-class). Pick a tier with `nvcl-ai install` or `nvcl-ai start --format`: see [Formats](/get-started/install.md#formats).

The four tiers apply to Qwen3.8-27B. DeepSeek-V4.1-Flash runs on DeepSeek's own 4- and 8-bit weights and MiMo-V2.6-Flash-RL on its checkpoint's own weights, both bit for bit as published (`exact`). GLM-5.3 runs on NVCL AI's own compressed weights for that model: `compact`, or Compact, a build of its own for two B200. Qwen3.8-Flash-Next runs on NVCL AI's own compressed weights for that model, about 7 bits per weight; its only format is also called `standard`, but it is a build of its own, not the Standard tier above.

* On NVIDIA GPUs, **Compact** has the fewest bits per weight, about 5.8. For one user it gives 7% more output than Standard (136 against 127 tokens/s; Qwen3.8-27B, one H200, drafts (speculative decoding) off, 1,024-token prompts and answers, greedy, same card and build, 23 October 2026).
* **Throughput** is opt-in, for many users at once on Hopper cards: why it writes more under load is on [Why it is fast](/concepts/why-fast.md#more-output-as-the-load-grows), results on [Two H200](/benchmarks/two-gpus.md#standard-and-throughput-on-two-cards).
* **Small** is the Mac default: 4-5 bits per weight, about 3.3x smaller than BF16. In its Mac build, Compact is about 2.6x smaller. What each needs: [Requirements](/get-started/requirements.md#mac).
* **Exact** stores the original BF16 weights in 12 bits instead of 16. No weight is rounded: each comes back as exactly its original value, so the weights you serve are the original, bit for bit.

These formats run only on NVCL AI, and we ship each model's weights ready to serve, with no conversion step on your side. Check them on your own evals: run Standard and Exact side by side.