Launching today

Gridline
One API to run batch inference, fine-tune and distill models
23 followers
One API to run batch inference, fine-tune and distill models
23 followers
Fine-tune, distill and batch open models at a fraction of the cost of closed models. Gridline follows the OpenAI API, so your existing code works with a new base URL. Train your own model and run it on real workloads, all from one API. Fine-tunes from $0.50, batch from $0.029 per 1M tokens, and $5 free to start.









Free Options
Launch Team

AWS Startup AdvisorPersonalized AWS guidance that meets you where you’re already building
Promoted
Hey Product Hunt, I'm Akshit, co-founder of Gridline.
Today we're launching Gridline: fine-tune, distill and batch open models at a fraction of the cost of closed models, through the OpenAI API you already use. Swap the base URL and key, and you're in.
We built it because having your own model meant either paying a lot or renting GPUs and wiring up training, evals and inference yourself. Builders should be building, not setting up infrastructure. Gridline closes the loop: train your model, check it improved, and run it at scale, all through one API.
What you get:
- Fine-tuning: upload examples, get a quote before it runs, see eval loss before and after, download your adapter. No GPU to pick, no training code.
- Batch inference on 13 open models (Qwen3.8 27B, Gemma 4, gpt-oss-20b…) or on your own fine-tune at the base model's price. No endpoint to host.
- One-call distillation: a big model teaches a small one.
A quick test we ran: Qwen3 8B went from 64.7% to 94.3% accuracy on Banking77 (PolyAI, CC BY 4.0) after a $0.50 fine-tune.
What Gridline did:
- Picked a gpu, trained, shut it down
- Turned the upload into a model in 27 min
- Scored base + fine-tune in one batch
- Charged $0.70, our two minimums (tokens alone: $0.42)
- Handed back the weights: an 87 mb lora file
Fine-tunes start at $0.50 a job, batch at $0.029 per 1M input tokens. As of today (19 Nov), we couldn't find a lower public price to fine-tune Llama 3.1 8B ($0.29/1M) or Qwen3 8B ($0.33/1M), or to batch Qwen3.8 27B, Gemma 4 26B and Qwen3.5 9B.
Our router sends each job to whichever GPU finishes it cheapest, including older cards big clouds don't bother with. That's how the prices stay this low.
New accounts get $5 free, no card.
What would you fine-tune first, and which models should we add next?
Docs: https://gridlinegpu.com/docs
Pricing: https://gridlinegpu.com/pricing
@akshit_pareek To answer your own question: I'd fine-tune a small model to read last night's GPU logs and sort out what actually broke, run as a batch every morning before my coffee.
Anyone else running infra, I'm here all day. Ask me anything about how the jobs actually run.