> 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/models.md).

# Supported models

The models NVCL AI runs today.

One NVCL AI server serves one model. Every model is served through the same OpenAI-compatible API; the request controls each one takes (reasoning effort, tool calls) are in [Per model](/reference/api.md#per-model).

| Model               | Status          | `nvcl-ai` name        | Formats (`--format`)                                                                                      |
| ------------------- | --------------- | --------------------- | --------------------------------------------------------------------------------------------------------- |
| Qwen3.8-27B         | Quality-checked | `qwen3.8-27b`         | `standard` (default), `compact`, `throughput`, `exact`; on a Mac `small` (default), `compact`, `standard` |
| DeepSeek-V4.1-Flash | Runs            | `deepseek-v4.1-flash` | `exact` (the published weights)                                                                           |
| GLM-5.3             | Runs            | `glm-5.3`             | `compact`, or Compact, a build of its own for two B200 (`nvcl-ai` 0.5.5 or newer)                         |
| Qwen3.8-Flash-Next  | Preview         | `qwen3.8-flash-next`  | `standard` (a build of its own)                                                                           |
| MiMo-V2.6-Flash-RL  | Runs            | `mimo-v2.6-flash`     | `exact` (the checkpoint's own weights)                                                                    |

Status: **Quality-checked** passed our task-level quality checks; **Runs** installs and serves like the others, without a task-level quality check yet; **Preview** is new in the catalog and still being checked.

On a Mac with Apple Silicon, Qwen3.8-27B runs as a preview; the other models need NVIDIA GPUs. `nvcl-ai models` shows which formats fit your machine. GPUs: [Hardware](/reference/hardware.md); what each model and format needs: [Requirements](/get-started/requirements.md).

## Qwen3.8-27B

|         |                                                                                                                                                                                                                                                                                                          |
| ------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Size    | 27B parameters, dense: every parameter is used for every token. A hybrid of linear-attention and attention layers, with a draft head (MTP) for speculative decoding.                                                                                                                                     |
| Context | The model's limit is 262,144 tokens. `nvcl-ai` serves it with a 128K-token context (32K a chat on a Mac). Through the chat template, Standard is level with vLLM 0.31 serving the full-precision model on RULER at 32K and 128K tokens (one H200): [Quality](/benchmarks/quality.md#long-context-ruler). |
| Inputs  | Text. Image input is planned.                                                                                                                                                                                                                                                                            |
| Weights | Four tiers, Standard by default; on a Mac three, Small by default. See [Weight formats](/reference/formats.md).                                                                                                                                                                                          |
| Quality | Standard: no measurable quality loss on our 1,015-task suite (measured on H200), run through the chat template. See [Quality](/benchmarks/quality.md).                                                                                                                                                   |

Total output against vLLM and SGLang: [One H200](/benchmarks/speed-h200.md).

## DeepSeek-V4.1-Flash

|         |                                                                                                                                        |
| ------- | -------------------------------------------------------------------------------------------------------------------------------------- |
| Size    | 552B parameters plus 196B of lookup tables. A mixture of experts: about 16B parameters are used for each token.                        |
| Context | The model's limit is 1,048,576 tokens. `nvcl-ai` serves it with a 128K-token context.                                                  |
| Inputs  | Text.                                                                                                                                  |
| Weights | DeepSeek's published weights, unchanged. The model was trained and published in 8 and 4 bits, so the published files are the original. |

Output on two H200: [Other models](/benchmarks/other-models-and-hardware.md).

## More models

| Model              | What it is                                                                                                                                                                                   |
| ------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| GLM-5.3            | 753B mixture of experts, about 40B active per token. Text only.                                                                                                                              |
| Qwen3.8-Flash-Next | 180B mixture of experts: 10 of its 512 experts work on each token. Text only. `nvcl-ai` serves it with a 160K-token context and answers up to 32K tokens.                                    |
| MiMo-V2.6-Flash-RL | About 310B mixture of experts, about 15B active. Text only, no tool calls. `nvcl-ai` serves it one chat at a time, with a 32K-token context: a second request waits until the first is done. |