Changelog
What's new
Product changes, newest first.
API tokens and an MCP server for coding agents
- Create personal API tokens under Settings. They are read-only unless you allow launching runs.
- The SynqTune MCP server lets Claude Code, Cursor and other agents check runs, read comparison reports, estimate costs and launch template pipelines.
- Launching from an agent always shows the plan and cost first, and starts only after you confirm.
Pipeline builder, task templates and comparison reports
- Pipelines are now built on a drag-and-drop canvas, with required steps locked in place.
- Templates for classification, extraction, tagging, redaction and summarization set the prompt format, training defaults and scoring.
- Compare a fine-tuned model with the LLM you use today on the same held-out examples, scored for the task, with a Ready / Close / Not ready verdict.
- Cost and return panel: GPU cost per hour and per day, tokens no longer sent to an API, and payback time.
- Copy any past run into the builder from the run's menu.
- Run pages show the pipeline as a graph with live step status.
Model versions and honest evaluation
- Models have numbered versions, and a new version can start from the settings of the previous one.
- Held-out examples are no longer part of training, and evaluation reports only measured values.
- Optional LLM grading scores and filters training data; it never rewrites responses.
More pipeline steps and chat with your model
- New data preparation steps: deduplication, quality and language filter, format conversion, augmentation and tokenizer training.
- Load a trained model on an instance in your account and chat with it.
Onboarding wizard
- Connecting an AWS account is a guided wizard with one CloudFormation stack.