Bring your JSONL, pick a base. Your first fine-tune runs in minutes.
Managed GPUs run the training loop, the scheduling and the checkpointing. Start from the dashboard with no code, or with one API call, and score the result against the base it started from.
- Base model
- Qwen3.8 27B
- Dataset
- tickets-2026.jsonl
One call replaces the loop
POST /v1/fine-tuning/jobs takes a base model and a training file, and the platform runs the training, the scheduling and the checkpointing.
Defaults you can leave alone
Epochs, learning rate, batch size and LoRA rank are all optional. Anything you omit is chosen for you, and the job reports the resolved values.
Watch the loss while it spends
The events stream carries the loss curve step by step and warns when validation loss starts to rise.
Cancel, and pay for what ran
Cancelling stops the run and stops the meter. billed_tokens_so_far is the total you are charged.
From a JSONL file to a scored checkpoint
Upload, choose, train, score. The platform validates the file, trains on managed GPUs and runs the evaluation you ask for. Take the dashboard with no code, the API with one call, or the training loop when the loss is yours to write.
Four steps to a scored checkpoint
Upload a JSONL file, or connect a source so the training set keeps itself current. Choose a base from the catalogue. Managed GPUs train it, with a checkpoint at every epoch. Then score the result against the base it started from.
Upload or connect
A JSONL of conversations. Validation rejects a bad row with its line number before any GPU spends.
Choose a base
Every trainable model carries a context window and a LoRA rank ceiling. The default is Qwen3.6 35B.
Managed GPUs train
The job moves from queued to running. Progress is a fraction, and the events stream carries the loss.
Score the checkpoint
An evaluation with compare_to turns one score into a before and after.
47 trainable bases, six modalities, three rank ceilings
47 of the 49 catalogue models accept a LoRA adapter. A fine-tuning job on a JSONL of conversations starts from one of the 26 text bases.
| Model | Context | LoRA rank |
|---|---|---|
| Rank 64 | ||
Nemotron 3.5 Lightningnvidia/Nemotron-3.5-Lightning | 131K | 64 |
GPT-OSS 20Bopenai/gpt-oss-20b | 131K | 64 |
Nemotron 3 Nano 30Bnvidia/Nemotron-3-Nano-30B-A3B | 66K | 64 |
Qwen3 8BQwen/Qwen3-8B | 33K | 64 |
Qwen3.8 27BQwen/Qwen3.8-27B-A3B | 262K | 64 |
Qwen3.6 35BQwen/Qwen3.6-35B-A3B | 262K | 64 |
| Rank 32 | ||
GLM 5.3 Flashzai-org/GLM-5.3-Flash | 1M | 32 |
GPT-OSS 120Bopenai/gpt-oss-120b | 131K | 32 |
Mistral Medium 3.5mistralai/Mistral-Medium-3.5 | 262K | 32 |
DeepSeek V4 Flash 0731deepseek-ai/DeepSeek-V4-Flash-0731 | 1M | 32 |
Llama 3.3 70Bmeta-llama/Llama-3.3-70B | 131K | 32 |
| Rank 16 | ||
MiniMax M3MiniMaxAI/MiniMax-M3 | 1M | 16 |
DeepSeek V3.2deepseek-ai/DeepSeek-V3.2 | 131K | 16 |
DeepSeek V3.1deepseek-ai/DeepSeek-V3.1 | 131K | 16 |
GLM 5.2zai-org/GLM-5.2 | 1M | 16 |
Kimi K2.6moonshotai/Kimi-K2.6 | 131K | 16 |
GLM 5.3zai-org/GLM-5.3 | 1M | 16 |
Nemotron 3 Ultra 550Bnvidia/Nemotron-3-Ultra-550B-A55B | 66K | 16 |
DeepSeek V4 Prodeepseek-ai/DeepSeek-V4-Pro | 1M | 16 |
Results from LLMs powered by Nucleus
monthly revenue misattributed
Ad Performance Tracking
trucks rerouted while still moving
Intelligent Fleet Optimisation
downtime prevented this month
Predictive Maintenance
yield shortfall traced to one cause
Agriculture Yield Analysis
Workflows driven by a fine-tuned model
Each of the Workflows runs on a custom model trained on Nucleus.
47 trucks rerouted while they were still moving
$73.4K a month was being credited to the wrong channel
Line 3's next failure, booked into the maintenance window
Fine-tuning at a glance
Data
- JSONL, one conversation per line, with system, user and assistant roles
- Validation before submit, with the rejected line and the reason
- A held-out validation file for val_loss and the overfit warning
- Connect a source and the training set keeps itself current
Bases
- 47 trainable models across six modalities
- A LoRA rank ceiling of 16, 32 or 64 per model
- Namespaced ids such as Qwen/Qwen3-8B
- Withdrawal dates published on the model ahead of time
Training
- sft and dpo
- Epochs, learning rate, batch size and rank, all optional
- Progress as a fraction and an events stream with the loss
- Cancel at any point and pay for the tokens that ran
After the run
- result_model as a nucleus:// path
- final_loss and val_loss on the job
- An evaluation with compare_to against the base
- A sampler on the OpenAI-compatible route, or export the weights
Training tokens by model group
Fine-tuning jobs and training-loop steps are metered per million tokens processed, whichever way you train.
| Group | Trainable bases | From | To |
|---|---|---|---|
| Language | 19 | $0.22 | $5.60 |
| Vision | 7 | $0.35 | $6.00 |
| Embedding & reranking | 6 | $0.03 | $0.09 |
| Image generation | 4 | $0.50 | $1.40 |
| Video generation | 3 | $1.40 | $1.80 |
| Audio | 8 | $0.28 | $0.90 |
USD per million training tokens. Mixture-of-experts models are priced by their active parameters.
More in Train
Training loop
DocsRL, DPO and distillation as API calls
Data connectors
Your systems as training data
Models
Open bases across six modalities
Read on
Bring your JSONL. The first fine-tune runs in minutes.
Managed GPUs, a scored checkpoint, and the weights are yours.