Yachay
Condor Models
Fine-tune open-source large language models on your own data — without a GPU, without a GCP account, without writing a training loop.
- 17
- base models
- LoRA
- & QLoRA
- $0.15/M
- tokens, production
- Managed
- GPUs
Curated catalog
17 commercial-safe base models — Llama 4 Scout, Llama 3.x, Qwen 3, Gemma 3, Mistral, Phi-4, DeepSeek Distill. License vetted; you keep the adapter.
LoRA or QLoRA
Parameter-efficient fine-tuning on managed cloud GPUs. We auto-pick L4, A100, or H100 based on your model — and the cheapest tune type that gets quality results.
From $0.15/M tokens
Cheaper than Fireworks and Together on equivalent workloads — production-size LoRA jobs typically land around $0.15-$0.20 per million training tokens vs their $0.48-$0.50. $5 minimum per job. No upfront bond — save a card on file once, we charge at completion.
No GCP account
No quotas, no Vertex billing setup, no IAM tickets. Sign in, pick a model, upload data, download an adapter.
How it works
From dataset to adapter in three steps
- 01
Pick a base model
Browse 17 commercial-safe open-source models — Llama 4, Llama 3.x, Qwen 3, Gemma 3, Mistral, Phi-4, DeepSeek Distill. We surface license, parameter count, and typical tune cost so the choice is informed before you commit.
- 02
Upload your data
Drop a JSONL, CSV, or TSV file — chat, instruction, or completion format, your choice. The submit form lints rows before you pay and shows an exact-price estimate based on parameters, dataset size, and epochs.
- 03
Download your adapter
We run the job on managed GPUs (L4 → A100 → H100, auto-sized). When it finishes, your adapter weights are delivered via a signed download URL. You keep the model; you keep the data.
Frequently asked
Do I need a GPU?
No. Yachay runs the job on managed Google Cloud GPUs — auto-sized to your model (L4 for small, A100 for mid, H100 for large). You upload a dataset, we run it, you download the resulting adapter.
Do I need a Google Cloud account?
No. Your Condor umbrella account covers billing — Yachay charges through the same Stripe customer you already use for Condor+, Vision, or Cybersecurity.
What dataset format do you accept?
JSONL (OpenAI chat, Alpaca, or ShareGPT) and CSV / TSV — whatever you've got. Yachay auto-detects and normalizes everything to OpenAI chat-style before fine-tuning. The submit form lints it before you pay; bad rows are surfaced inline.
Who owns the tuned model?
You do. We claim no IP interest in your dataset or the resulting adapter. Use is subject to the upstream base model's license, which we summarize on the catalog page.
How fast does a tune finish?
A typical 8B-parameter LoRA over a 20 MB dataset finishes in 15–40 minutes. 70B jobs are a few hours. The estimator shows the projected duration before you commit.
What happens if my job fails?
Infrastructure failures on our side never fire an invoice. Since billing is charge-at-completion against your saved card, "no invoice" literally means "no charge" — no refund timing to worry about. If a job somehow did clear and is subsequently marked failed for an infrastructure reason, we refund automatically via Stripe and the dashboard shows the refund confirmation. Dataset issues that slipped past the linter are reviewed case by case at hello@condorbox.ai.
Why the name
Yachay is Quechua for knowledge — and for the act of learning itself. The product is named for what it does: turn your data into a model that knows.