> ## Documentation Index
> Fetch the complete documentation index at: https://docs.platform.toloka.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Base models

> Reference the Training model catalog, method compatibility, GPU counts, pinned revisions, and model lineage.

A Base model is the catalog model a Training starts from. Every Training Run uses a pinned upstream snapshot, and every published Model Version records the Base model it needs.

## Catalog

| ID | Display name | Provider | GPUs per Training Run | SFT | Gisting | Pinned snapshot |
| - | - | - | -: | - | - | - |
| `qwen3-4b` | Qwen3 4B | Qwen | 1 | Supported | Supported | [Qwen/Qwen3-4B-Instruct-2507 @ cdbee75f](https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507/tree/cdbee75f17c01a7cc42f958dc650907174af0554) |
| `qwen3-8b` | Qwen3 8B | Qwen | 1 | Supported | Supported | [Qwen/Qwen3-8B @ b968826d](https://huggingface.co/Qwen/Qwen3-8B/tree/b968826d9c46dd6066d109eabc6255188de91218) |
| `qwen3-32b` | Qwen3 32B | Qwen | 8 | Supported | Not supported | [Qwen/Qwen3-32B @ 9216db57](https://huggingface.co/Qwen/Qwen3-32B/tree/9216db5781bf21249d130ec9da846c4624c16137) |
| `qwen3-235b` | Qwen3 235B | Qwen | 8 | Supported | Not supported | [Qwen/Qwen3-235B-A22B-Instruct-2507 @ ac9c66cc](https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507/tree/ac9c66cc9b46af7306746a9250f23d47083d689e) |

List the current catalog from the CLI:

```sh theme={null}
toloka base-model list
```

Inspect one entry:

```sh theme={null}
toloka base-model get qwen3-8b
```

Both commands show the upstream repository and pinned revision.

## Method compatibility

SFT publishes a LoRA adapter, so all four catalog entries are supported.

Gisting publishes full weights. Only `qwen3-4b` and `qwen3-8b` fit within Model Version size limits. Creating a gisting Training on a larger tier is rejected with:

```text theme={null}
Gisting publishes a full model and supports the qwen3-4b and qwen3-8b Base Models; larger tiers exceed the Model Version size limits
```

The eight-GPU SFT models process about eight times as many estimated tokens within the six-hour wall-time limit. For the same token count, their maximum charge is about the same as a one-GPU model except where the per-GPU minimum charge dominates.

## Model sources

A new Training can start from either:

* A catalog Base model, passed as `--base-model <id>`.
* A LoRA adapter Model Version in the same project, passed as `--model-version <id>`.

Starting from a Model Version continues from its adapter weights. SFT continues training the adapter. Gisting merges the adapter into its Base model before learning gist tokens.

A full-weight Model Version cannot be a Training source. The platform rejects it with:

```text theme={null}
Training from a Model Version requires a LoRA adapter source
```

After the first Training Run creates the output Model, changing the source is allowed only when the replacement has the same Base model.

## Pinned revisions and lineage

Each Training Run resolves a pinned snapshot from the Base model catalog. A Model Version records its Base model ID, content kind, file manifest, and training parent. It does not persist the upstream repository or revision; the catalog carries the current values.

An adapter's `adapter_config.json` may omit the revision, so the file may not identify the trained snapshot.

Before serving a downloaded adapter, inspect its configuration and the Base model catalog. If the adapter has no revision and the catalog may have changed since training, confirm the compatible snapshot rather than assuming the current catalog reconstructs its history. The platform does not host or serve the adapter or full model.

## Output shape

An SFT Model Version contains `adapter_model.safetensors`, `adapter_config.json`, and any additional validated `.json` files from the run.

A gisting Model Version includes `config.json` and either `model.safetensors` or a complete numbered set of `model-*.safetensors` weight shards. Tokenizer files, `merges.txt`, and `chat_template.jinja` are optional sidecars, not guaranteed contents.

The learned prompt is represented by `<|gist_…|>` tokens in that full Model Version.

Model Version file and download constraints are listed in [Training limits](https://docs.platform.toloka.ai/docs/train/reference/limits.md#model-version-limits).

## Related

* [Choose a training method](https://docs.platform.toloka.ai/docs/train/get-started/choose-method.md)
* [Training settings](https://docs.platform.toloka.ai/docs/train/reference/settings.md)
* [Training cost](https://docs.platform.toloka.ai/docs/train/concepts/cost.md)


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.