Models

Catalog

Use the id as model when you create a fine-tuning job, or as body.model in a batch. GET /openai/v1/models lists the models your account can use right now.

• The chat models take up to 8,192 tokens of context in a batch and write at most 1,024 tokens per request; gpt-oss-20b takes up to 16,384 and writes at most 4,096.
• gpt-oss-20b reasons before it answers. The reasoning is written at low effort, counts against max_tokens and is billed as completion tokens; the result holds the answer alone. If the reasoning uses the whole allowance the answer comes back empty, so leave max_tokens room for both. It does not take a JSON response_format.
• gpt-oss-20b, qwen3.5-9b and qwen3.8-27b take a system message only as the first message, and need at least one user message. qwen3.5-9b and qwen3.8-27b are served for text only.
• nemotron-3.5-lightning-30b-a3b is NVIDIA's own 4-bit (NVFP4) build of the model, under the OpenMDW-1.1 licence.
• qwen3-embedding-0.6b returns dense embeddings, 1,024 dimensions, normalized, for inputs of up to 8,192 tokens. It embeds exactly the text you send: for retrieval, put the instruction the model card describes in front of each query yourself and send documents bare.
• bge-m3 returns dense embeddings only, 1,024 dimensions, for inputs of up to 8,192 tokens. It is served from a safetensors conversion of BAAI's commit above.
• llama-3.1-8b-instruct is Meta's Llama 3.1 8B Instruct, served from Red Hat's ungated copy of Meta's files (the same bytes). Built with Llama: its use is subject to the Llama 3.1 Community License and Acceptable Use Policy.
• qwen3-8b answers with thinking off in batches, the form its fine-tunes are trained in.
• A qwen3-8b fine-tune made before batch serving of qwen3-8b began kept its LoRA adapter for download and created no fine_tuned_model: that job's fine_tuned_model is null and gridline.fine_tuned_adapter names the adapter (servable false). A new job on the same files creates a fine-tuned model you can run batches on.

Gridline is an independent service. Qwen, Mistral, Llama, gpt-oss, Nemotron, Gemma and bge-m3 are open models from their respective developers; Gridline is not affiliated with or endorsed by them.

Your fine-tuned models

A fine-tuning job that succeeds registers its adapter under an ft: id, for example:

ft:qwen2.5-7b-instruct:o3f9a1c2d4e5:7b2c9e41a0d3

GET /openai/v1/models lists your ft: models alongside the catalog. No other account sees them, and naming one from another account is answered model_not_found, as is an id that does not exist. The fine-tuning reference explains each part of the id, how to download the adapter, and how to delete the model. An account holds at most 100 fine-tuned models.

How fine-tuned models are served

• A batch on an ft: model runs its base model at the exact commit the adapter was trained on, with the adapter loaded.
• One base model per batch, mixed with up to 8 ft: models of that base, within what the base has been proved to serve: on qwen2.5-0.5b-instruct, qwen2.5-1.5b-instruct, 8 of any rank; on qwen2.5-7b-instruct, mistral-7b-instruct-v0.3, llama-3.1-8b-instruct, 4 adapters of rank 64 or 8 of rank 32 or lower; on qwen3-8b, 4 of any rank.
• Only bases in the batch catalog can serve batches. A model fine-tuned on any other base stays downloadable but cannot be named in a batch.
• A QLoRA adapter is trained against a 4-bit quantized base and served on the bf16 base, so its outputs can differ slightly from what training measured.

Not offered

• Qwen2.5-3B: its licence is not Apache-2.0.
• Models of 70B parameters and larger.
• Fine-tuning embedding models.