Neptune
Experiment tracking and model observability platform built for large scale training with high throughput logging dashboards alerts and enterprise controls.
Neptune has been discontinued. The product was shut down, absorbed into another company, or now operates under a different brand. The information below is kept for reference and may describe the tool as it was.
Overview
Neptune centralizes experiment metadata metrics and artifacts so ML teams can debug training faster and keep models reproducible. Engineers log losses gradients activations and custom metadata from any framework then slice runs by tags params or code commits. Dashboards render charts instantly even at high cardinality while alerts flag regressions and stalled jobs.
Artifact management stores checkpoints predictions and datasets with versioning and lineage to the exact code and data used. Role based access SSO and audit logs meet enterprise requirements and ingestion SLAs ensure stability during big runs. Teams adopt Neptune to reduce wasted GPU cycles shorten iteration loops and create a durable history of how models evolved from baseline to production candidates.
Self host and cloud options exist with published Startup and Lab tiers for foundation model scale and API compatibility across Python and popular orchestration stacks.
Key features
- High throughput logging: Capture millions of metrics with no missed spikes during large scale training
- Artifacts and lineage: Store checkpoints datasets and predictions with code and data version links
- Fast dashboards: Slice compare and overlay runs with tags params and commits at interactive speed
- Alerts and regressions: Detect stalled jobs metric drops and drift with notifications to chat and email
- Role based access: Enforce SSO RBAC and audit logs for enterprise teams and compliance
- APIs and SDKs: Integrate with PyTorch TensorFlow and orchestration tools quickly
- Ingestion SLA: Rely on published uptime and data guarantees for stability at peak loads
- Cloud or self host: Choose managed service or private deployment to meet security needs
Best for
- Track and compare baselines and ablations across teams
- Debug exploding loss or instability with fine grained metrics
- Version artifacts and link to exact code and data
- Share dashboards for reviews and model sign offs
- Alert on regression after code or data changes
- Create reproducible histories for audits and handoffs
- Shorten iteration loops during foundation model training
- Migrate from ad hoc notebooks to governed workflows
Capabilities
Metrics at scale
Stream per step metrics and custom signals then query interactively without losing fidelity or spikes.
Artifacts and lineage
Capture checkpoints datasets and predictions with links to commits params and environment fingerprints.
Dashboards and alerts
Create comparisons share views and receive notifications when metrics drift or jobs stall.
Roles and audit
Use SSO RBAC and audit logs and deploy cloud or self hosted to meet org security policies.
Frequently Asked Questions
How does Neptune pricing start?
Public pricing lists Startup at $150 per user per month and Lab at $250 per user per month with trials available.
Is there a free option?
Trials are available and historical community tiers existed but current focus is paid team plans check the pricing page for updates.
Can we self host Neptune?
Yes self hosted deployments are supported for customers needing full control of data and networking.
Which frameworks are supported?
Integrations cover PyTorch TensorFlow scikit learn and common orchestration tools with simple client APIs.
How do you handle very large runs?
High throughput ingestion and SLAs plus artifact deduplication keep logging stable at foundation model scales.
