Gridline
Intro
Gridline is a fine-tuning and batch-inference API launched on 18 November 2026. The official site gridlinegpu.com describes it as one API for fine-tuning, distillation, and batch inference on open models: you keep the weights, and Gridline runs the GPUs. The homepage and llms.txt both say text fine-tuning and batch inference are live today, on an OpenAI-compatible API.
The same day, @apareek05 posted “Today we're launching Gridline.” (https://x.com/apareek05/status/2123055594504687700), and a Show HN thread introduced the service. The Quickstart shows how to point an OpenAI client at https://api.gridlinegpu.com/openai/v1 with a dashboard-issued GRIDLINE_TOKEN, then upload a training file, create a fine-tune, or submit a batch.
This entry records the product name Gridline for this API service. It should not be merged with unrelated companies or projects that share the same everyday word.
Advertisement
Explain
Gridline puts three jobs behind one OpenAI-compatible API. Fine-tuning: upload at least 11 examples as OpenAI chat JSONL; the job holds some rows out for evaluation and needs 10 to train on. It trains a bf16 LoRA adapter by default, and uses QLoRA on smaller bases. A successful model is named like ft:qwen3.5-9b:..., and you can download adapter_model.safetensors and adapter_config.json with your token. Batch inference: one request per line, with a 24-hour or 72-hour completion window (the docs say the price is the same), and results come back under your custom_id. Embeddings use /v1/embeddings with bge-m3. Distillation: run a batch on a large teacher, call POST /files/distill to turn the results into training and validation files, then fine-tune a smaller student. The docs say that conversion call is not charged on its own.
Billing uses prepaid credits. A fine-tune is quoted when you create it, and the charge after success is never more than the quote; creating the job starts it. A batch holds credits when it is validated and settles when it ends. If credits cannot cover the work, the job fails with insufficient_quota before anything runs. An account trains one fine-tune at a time, and can have 16 fine-tunes and 64 batches open at once. The docs also say a verified signup includes $5 of free credit while that day’s credits last, and that a dashboard token carries the read and routing:write scopes, with a 30-, 60-, or 90-day expiry.
The company also draws a boundary: speech and vision models, and more bases, are described as next. The site says Gridline is an independent service and is not affiliated with or endorsed by the developers of Qwen, Mistral, Llama, gpt-oss, Nemotron, Gemma, or bge-m3. It is backed by Entrepreneurs First and is a newly launched self-serve API, so evaluate it from the live docs and what you can actually run, not as a long-established cloud vendor.
Usage
Sign up at gridlinegpu.com, verify your email, and create a token from the dashboard as shown in the Quickstart. Then set:
export GRIDLINE_TOKEN="gla_..."
Point the OpenAI Python SDK at https://api.gridlinegpu.com/openai/v1, upload a JSONL file with purpose=fine-tune to start a job, or upload a purpose=batch file for batch inference. The docs use catalog models such as qwen3.5-9b or bge-m3. After a fine-tune succeeds, set body.model in the batch file to your ft: model. The docs say one batch can use up to eight of your ft: models that share a base, and that other accounts cannot call your models.
For distillation, follow Distillation: run a teacher batch, call /files/distill on the output, then create a fine-tune on the student. Read job.gridline["quote"] before you rely on the job, and make sure your credits cover it. You can also install the CLI with python -m pip install 'gridlinegpu[source-contract]>=0.3.2,<0.4' and run gridline setup --apply.
If you only want to fine-tune on your own GPUs, you do not need this new hosted API name. Before you upload data or teacher outputs, confirm you have the rights to train on them.