A friend who works on recommendation algorithms told me last week that his team’s experiment records had turned into a mess. Three people were each running their own training scripts, hyperparameters were scattered across a spreadsheet, and model checkpoints piled up in different folders on a shared drive. Nobody could say for certain which training run produced the model currently in production, or what data it had been trained on. It took a production incident, one they couldn’t trace back to a specific configuration, before the team finally decided to adopt a proper experiment tracking tool.
This scenario is common enough to be almost universal. Writing model code isn’t the hard part. Remembering exactly what you ran is. Parameters get tweaked a dozen times, and nobody can reliably recall which version performed best or which data snapshot was used. A category of tools exists specifically to solve this: log every run’s parameters, metrics, and output artifacts so teams can compare and reproduce experiments without relying on memory.
The answer to “which tool should I use” has shifted more than once over the past few years. A handful of events in early 2026 reshaped this space enough that they’re worth covering before getting into the tools themselves.
The category had a rough year
Neptune.ai used to be a well-regarded option in this space, built specifically for large-scale foundation model training, the kind of workload that runs for months and generates enormous volumes of experiment data. This year it was acquired by OpenAI. The official announcement was blunt: the hosted service will be permanently shut down following a three-month transition period, ending March 5, 2026. After that date, any remaining data gets deleted and there’s no recovering it. Anyone still on Neptune needs to export their data and find somewhere else to land.
Weights & Biases (commonly shortened to W&B) went through something similar, acquired by CoreWeave. For day-to-day users, nothing changes immediately, the interface and features stay the same. But if you’re about to sign a multi-year enterprise contract, it’s worth confirming directly whether pricing or deployment terms might shift under the new ownership.
A few months before that, DVC, the data version control project, was acquired by lakeFS. DVC itself remains open source and actively maintained. But the acquisition makes a point worth sitting with: even when a license says Apache 2.0, the company behind a project can disappear overnight, and you find out through a press release after the fact.
Put these three events together and they point at the same underlying question: if the tool you picked is backed by a venture-funded startup, that company can get acquired, shut down, or pivot its business model at any point. Evaluating a tool on features and pricing alone isn’t enough anymore. Governance structure matters just as much.
This is part of why MLflow has quietly become something like the “safe default” in this category. It was donated to the Linux Foundation back in 2020, which means governance doesn’t sit with any single company. You might not love its interface, but you’re not going to wake up one morning to an email announcing the service shuts down in three months.
What MLflow actually offers
Worth being precise about what MLflow provides. It started as a Databricks project and was later donated to the Linux Foundation, released under Apache 2.0, entirely free, with no separate paid tier gating features behind a paywall. Experiment tracking, model registry, evaluation, tracing for LLM-based applications, prompt management, all of it ships in the open source release with nothing held back.
Deployment is straightforward: stand up a service, connect a database and object storage, and you’re running. The official repository ships a working Docker Compose setup, and adjusting a few environment variables gets a full instance with PostgreSQL and a storage backend running within minutes. If you’re already on Databricks, AWS, Azure, or GCP, each of those platforms offers a managed version of MLflow too, so teams that don’t want to run their own servers can offload the operational side while keeping the open source data format, avoiding vendor lock-in.
The tradeoffs are real. The interface is functional rather than polished, and comparing large numbers of runs side by side is slower than in some of the commercial tools covered below. Self-hosting also means someone on the team owns the database, storage, and upgrade path, and that operational overhead isn’t free even when the software is.
What it costs to buy convenience with W&B
Weights & Biases takes a different approach entirely: a polished interface, solid team collaboration features, and hyperparameter search (Sweeps) that tends to work more smoothly than most competitors. A lot of research teams choose it specifically for how quickly people can start using it without any setup friction.
That convenience shows up on the invoice. The free tier covers 5 projects and 1 team member, really only enough for individual experimentation. Actual team collaboration requires the Pro tier, which starts around $60 per month depending on team size, while enterprise pricing is negotiated directly and varies widely, anywhere from a few hundred to several thousand dollars a month depending on the deal.
One rough estimate circulating online put self-hosted MLflow costs at under $100 a month for a modest setup, a small VM plus a managed database and object storage. What that estimate leaves out is the engineering time required to actually run it: patching the server, managing the database, handling backups. None of that is free. MLflow trades subscription fees for a team’s own operational time. W&B does the opposite, the subscription buys you out of dealing with any of that. The math only really matters once a team is buying more than one or two seats.
ClearML: folding the whole pipeline into one system
If MLflow focuses narrowly on tracking experiments, ClearML aims at something bigger, folding experiment tracking, task scheduling, data versioning, model repository, and pipeline orchestration into a single system, closer to a full MLOps stack than a point solution.
ClearML’s open source core is also Apache 2.0. The free Community edition is self-hostable with no user limits and includes the core tracking, model repository, and pipeline features. There’s also a free hosted server option, capped at 3 users and certain usage limits. Beyond that, the Pro tier starts at $15 per user per month, plus usage-based charges for storage and API calls, aimed at teams that need finer-grained access control and automation as they scale.
Teams that pick ClearML usually want one system managing everything rather than stitching together separate tools for tracking, scheduling, and data versioning. The tradeoff is a steeper learning curve. If all you need is basic experiment logging, ClearML brings a lot of machinery you may never touch.
Comet: from experiment tracking to LLM observability
Comet’s focus has visibly shifted toward large language models over the past couple of years. It open-sourced a sub-project called Opik specifically for LLM application tracing and evaluation, also under Apache 2.0, with the self-hosted version carrying the full feature set, nothing held back for a paid tier. As of early September this year, the project had crossed twenty thousand stars on GitHub, growing at a decent clip.
Pricing follows the same structure as Comet’s main platform: a free tier with 25,000 tracked spans per month, 60-day data retention, and up to 10 team members, followed by a Pro tier starting at $19 per month that raises the cap to 100,000 spans and 50 team members. The enterprise tier adds single sign-on, access controls, and a set of compliance certifications, priced through direct negotiation.
The division of labor between Comet and MLflow is fairly clean. For traditional model training experiments, MLflow or the other tools covered here are generally sufficient. For teams building LLM applications or agents that need to trace every step of a call chain, tools purpose-built for LLM observability, Comet’s Opik among them, save considerable time compared to assembling an observability stack from scratch.
Putting the options side by side
| Tool | License | Self-hosted | Free tier limits | Paid tier starts at | Primary focus |
|---|---|---|---|---|---|
| MLflow | Apache 2.0 (Linux Foundation) | Default deployment, fully free | No limits (no feature gate when self-hosted) | No paid tier; managed hosting billed separately | Full lifecycle, most stable governance |
| Weights & Biases | Closed-source SaaS | Supported, not the focus | 5 projects / 1 member | ~$60/month | Team collaboration and visualization |
| ClearML | Apache 2.0 core + paid tiers | Yes, free Community edition | 3 users (hosted) | $15/user/month | Full pipeline in one system |
| Comet (Opik) | Apache 2.0 | Yes, full feature set | 25,000 spans/month, 10 members | $19/month | LLM observability |
| DVC | Apache 2.0 (now under lakeFS) | Yes | No limits | Free | Data version control, not experiment tracking |
DVC deserves a note here because it isn’t really the same category of tool. It versions data and model files, essentially Git for data, which makes it complementary to the tools above rather than a substitute. Plenty of teams run MLflow for experiment tracking and DVC for data versioning side by side, not as an either-or choice.
So which one makes sense
Back to that friend’s situation. If a team already has someone who can run a server and doesn’t want to pay a subscription for the tool itself, MLflow is the default choice, and its governance stability carries more weight than it used to, especially after this year’s string of acquisitions and shutdowns. If a team values getting started without operational overhead and the budget allows for per-seat pricing, W&B’s visualization and collaboration experience is still the most polished in the category, though it’s worth keeping an eye on how CoreWeave’s ownership shapes long-term pricing.
For teams managing more than just model training, data pipelines, task scheduling, multi-user collaboration, ClearML’s all-in-one approach is worth a look. For teams whose work has shifted toward LLM applications and agents, where the thing that needs tracking is an entire call chain rather than a handful of training metrics, Comet’s Opik fits that use case more directly.
Neptune, at this point, isn’t really an option to consider for anything new. Its hosted service is winding down, and building on it now just means migrating again in a few months. That’s probably the most direct takeaway from this whole roundup: who maintains a tool, how that company makes money, and how stable its governance structure is matter as much as the feature list. Tools evolve constantly, but nobody wants to be told, halfway through a project, that they have three months to move somewhere else.
Related reading
- When Composio Stops Being Enough: Nango, Arcade, or Pipedream Connect
- Dragonfly vs Redis: Which In-Memory Database Should You Pick
- How to Choose a TypeScript AI Agent Framework: Mastra and Its Alternatives
- Plausible vs Umami vs Fathom: Choosing a Privacy-First Web Analytics Tool
- Qwen Code and Open-Source Terminal AI Coding Assistants: A 2026 Comparison
Stay updated with our latest AI insights
