Supported base models

Hugging Face base models Loops accepts, with sequence-length limits and vision support.

Loops supports the base models below for LoRA training. Pass the Hugging Face repo ID as base_model when starting a run.

Every Baseten workspace can train models up to Qwen3.8-27B, which supports sequences up to 262,144 tokens. To train larger models or run more training sessions at once, request more access.

Models

Model	Max sequence length	Vision
Qwen/Qwen3-0.6B	8,192	No
Qwen/Qwen3-8B	40,960	No
Qwen/Qwen3.5-0.8B	131,072	Yes
Qwen/Qwen3.5-2B	131,072	Yes
Qwen/Qwen3.5-4B	131,072	Yes
Qwen/Qwen3.5-9B	131,072	Yes
Qwen/Qwen3.5-27B	131,072	Yes
Qwen/Qwen3.5-35B-A3B	131,072	Yes
Qwen/Qwen3.5-122B-A10B	262,144	Yes
Qwen/Qwen3.5-397B-A17B	262,144	Yes
Qwen/Qwen3.6-27B	131,072	Yes
Qwen/Qwen3.6-35B-A3B	262,144	Yes
Qwen/Qwen3.8-27B	262,144	Yes
deepseek-ai/DeepSeek-V4-Flash	131,072	No
deepseek-ai/DeepSeek-V4-Flash-0731	131,072	No
google/gemma-4-26B-A4B	65,536	No
google/gemma-4-26B-A4B-it	65,536	No
google/gemma-4-31B	131,072	No
google/gemma-4-31B-it	131,072	No
moonshotai/Kimi-K2.6	262,144	Yes
moonshotai/Kimi-K2.7-Code	262,144	Yes
moonshotai/Kimi-K3	262,144 (SFT), 131,072 (RL)	No
nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16	262,144	No
nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16	262,144	Yes
nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4	262,144	Yes
nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4	262,144	No
zai-org/GLM-5.2-FP8	262,144	No
zai-org/GLM-5.3	262,144	No
zai-org/GLM-5.3-Flash	262,144	Yes

Some models restrict the available loss functions by configuration. GLM 5.3 runs long contexts with context parallelism and router replay (R3) enabled, and in that configuration cross_entropy and dpo are unavailable. See Loss functions for the full list.

List supported models

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

curl --fail-with-body --get \
  --url https://api.baseten.co/v1/loops/capabilities \
  --header "Authorization: Bearer $BASETEN_API_KEY" \
  --data-urlencode "use_case=sft"

The response includes each model’s ID, supported sequence length (max_seq_len), and sequence length enabled for your workspace (max_enabled_seq_len). Use use_case=rl to list models for reinforcement learning.

Pass a model to Loops

Pass the table value verbatim as base_model through any of the following entry points:

• The Python SDK, using ServiceClient.create_lora_training_client(base_model=...). See the Loops quickstart.
• The HTTP API, using POST /v1/loops/trainers.
• The CLI, using baseten loops run create --base-model <base_model>, which provisions a session, run, and paired sampler in one call.

The minimal HTTP call provisions a trainer against an existing session. Replace 2qjl22w with the session.id returned by POST /v1/loops/sessions:

curl --request POST \
  --url https://api.baseten.co/v1/loops/trainers \
  --header "Authorization: Bearer $BASETEN_API_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "session_id": "2qjl22w",
    "base_model": "Qwen/Qwen3.5-9B"
  }'

For the full request body and response shape, see the Loops API reference.

Request more access

Use the Loops request form for larger models, longer sequences, more concurrent runs, or a model that isn’t listed. Include:

• Your workspace name and the model’s Hugging Face repo ID.
• What you’re training and whether you need supervised fine-tuning or reinforcement learning.
• The maximum sequence length and number of concurrent runs you need.
• When you need the capacity.