Supported models

Choose a base model for supervised fine-tuning (SFT) or reinforcement learning (RL). Every model trains in the dedicated mode, and three also train in the serverless mode. To compare the modes, see Availability modes.

Choose a model

To pay per token without managing deployments, choose a serverless model:

Qwen/Qwen3.6-35B-A3B: text and vision.
zai-org/GLM-5.3-Flash: text and vision.
zai-org/GLM-5.3: text only. Its long-context configuration doesn’t support cross_entropy or dpo losses.

Serverless runs bill per token at the rates listed under Serverless Loops token pricing. To train any other model, or to choose GPUs and replicas, use the dedicated mode.

Models

Model	Availability	Max sequence length	Vision
`Qwen/Qwen3.6-35B-A3B`	Serverless and dedicated	262,144	Supported
`zai-org/GLM-5.3`	Serverless and dedicated	262,144	Not supported
`zai-org/GLM-5.3-Flash`	Serverless and dedicated	262,144	Supported
`Qwen/Qwen3-0.6B`	Dedicated	8,192	Not supported
`Qwen/Qwen3-8B`	Dedicated	40,960	Not supported
`Qwen/Qwen3.5-0.8B`	Dedicated	131,072	Supported
`Qwen/Qwen3.5-2B`	Dedicated	131,072	Supported
`Qwen/Qwen3.5-4B`	Dedicated	131,072	Supported
`Qwen/Qwen3.5-9B`	Dedicated	131,072	Supported
`Qwen/Qwen3.5-27B`	Dedicated	131,072	Supported
`Qwen/Qwen3.5-35B-A3B`	Dedicated	131,072	Supported
`Qwen/Qwen3.5-122B-A10B`	Dedicated	262,144	Supported
`Qwen/Qwen3.5-397B-A17B`	Dedicated	262,144	Supported
`Qwen/Qwen3.6-27B`	Dedicated	131,072	Supported
`Qwen/Qwen3.8-27B`	Dedicated	262,144	Supported
`deepseek-ai/DeepSeek-V4-Flash`	Dedicated	131,072	Not supported
`deepseek-ai/DeepSeek-V4-Flash-0731`	Dedicated	131,072	Not supported
`google/gemma-4-26B-A4B`	Dedicated	65,536	Not supported
`google/gemma-4-26B-A4B-it`	Dedicated	65,536	Not supported
`google/gemma-4-31B`	Dedicated	131,072	Not supported
`google/gemma-4-31B-it`	Dedicated	131,072	Not supported
`meta-models/Muse-Spark-Open-816B-A42B`	Dedicated	262,144 (SFT only)	Not supported
`meta-models/Muse-Spark-Open-816B-A42B-NVFP4`	Dedicated	262,144	Supported
`moonshotai/Kimi-K2.6`	Dedicated	262,144	Supported
`moonshotai/Kimi-K2.7-Code`	Dedicated	262,144	Supported
`moonshotai/Kimi-K3`	Dedicated	262,144 (SFT), 131,072 (RL)	Not supported
`nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16`	Dedicated	262,144	Not supported
`nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16`	Dedicated	262,144	Supported
`nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4`	Dedicated	262,144	Supported
`nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4`	Dedicated	262,144	Not supported
`zai-org/GLM-5.2-FP8`	Dedicated	262,144	Not supported

Check workspace access

Use Get server capabilities to list the models and sequence lengths your workspace can use. Set your workspace API key as BASETEN_API_KEY:

export BASETEN_API_KEY="<YOUR_API_KEY>"

Baseten CLI
REST API

To check model access:Install the Baseten CLI. For reinforcement learning (RL), run the following command:

baseten api management 'loops/capabilities?use_case=rl'

For supervised fine-tuning (SFT), run the following command:

baseten api management 'loops/capabilities?use_case=sft'

To check model access:For reinforcement learning (RL), run the following command:

curl --url 'https://api.baseten.co/v1/loops/capabilities?use_case=rl' \
  --header "Authorization: Bearer $BASETEN_API_KEY"

For supervised fine-tuning (SFT), run the following command:

curl --url 'https://api.baseten.co/v1/loops/capabilities?use_case=sft' \
  --header "Authorization: Bearer $BASETEN_API_KEY"

Both interfaces return a supported_models list. This excerpt shows one model; the available models and limits depend on your workspace:

Response excerpt

{
  "supported_models": [
    {
      "model_name": "Qwen/Qwen3.5-2B",
      "max_seq_len": 131072,
      "max_enabled_seq_len": 131072,
      "max_context_length": 131072,
      "supports_vision_language": true,
      "enabled": true,
      "enablement_details": null
    }
  ]
}

Each model entry includes:

string

Model ID to pass as base_model when you create a run.

integer

Maximum supported training sequence length, in tokens.

integer

Maximum training sequence length enabled for your workspace, in tokens. A value of 0 means your workspace can’t run the model.

boolean | null

Whether serverless Loops supports the model. Support doesn’t guarantee that a particular run or sampler uses the serverless mode. See Availability modes for how Loops chooses a mode.

An enabled model can still fail to provision when GPUs are unavailable or your organization has reached its trainer limit. See Resolve capacity and quota errors to choose whether to retry, stop an existing run, or request more capacity.

Next steps

Loops quickstart: Train and sample with serverless Loops.
Choose a base model: Compare candidates against your task’s constraints and evaluation.
Request more access: Ask for additional models, longer sequences, or more concurrent runs.

Get started

Model APIs

Baseten Grid

Inference

Frontier Gateway

Training

Sandboxes

Development

Deployment

Engines

Model Labs

Organization

Hosting options

Observability

Troubleshooting

Supported models

Choose a model

Models

Check workspace access

Next steps

​Choose a model

​Models

​Check workspace access

​Next steps

Choose a model

Models

Check workspace access

Next steps