"MLOps" gets thrown around as one big category, but in practice it's several different jobs: getting the right data to your model, finding somewhere cheap to train it, getting it into production without a mess, storing the embeddings it produces, checking whether it's actually getting better or worse, and keeping track of what it all costs. Most teams stitch together 3-4 separate tools to cover this instead of buying one bloated "platform". Here are 8 real tools — one or two per job — pulled from tools actually listed in our catalog, with honest pricing and trade-offs so you can see where each one fits.
Feature stores: getting consistent data to your models
A feature store keeps the data you used to train a model in sync with the data your model sees in production — without it, you get subtle bugs where a model performs great in testing and then falls apart live because the two data paths drifted apart.
Feast — the free, DIY option
Feast is a fully open-source (Apache-2.0) feature store with no paid tier at all. It plugs into 10+ offline and 15+ online stores you probably already run, and does point-in-time-correct joins to prevent training-serving skew. The catch: there's no managed hosting or commercial support baked in, so you need in-house data engineering to run it at scale. Pick Feast if you have the engineering capacity to self-host and want zero vendor lock-in.
Hopsworks — managed, and more than just a feature store
Hopsworks bundles a feature store, an AI lakehouse and MLOps tooling into one platform, with sub-millisecond online feature serving and support for Delta, Iceberg and Hudi table formats. It has a genuinely free tier (1 project, no credit card), then pay-as-you-go SaaS or custom enterprise pricing with on-premise deployment. The trade-off: it's licensed AGPL-3.0 (stricter copyleft than Feast's Apache-2.0) and is a heavier, more opinionated platform than a single-purpose feature store. Pick Hopsworks if you want a managed option and are fine consolidating more of your stack into one vendor.
Training compute: finding and paying for GPUs
Modal — serverless Python on GPUs
Modal runs your Python code — including training and inference — on GPUs with sub-second cold starts, and you define infrastructure as plain Python instead of YAML. Billing is per-second with zero idle charges: GPU compute starts at $0.000164/sec. There's a free Starter tier with $30/month in compute credits, a Team plan at $250/month (plus compute, with $100/month in credits), and custom Enterprise pricing. The catch: the platform itself is proprietary (only the client SDK is open source), and always-on high-volume inference can end up cheaper on reserved instances elsewhere. Pick Modal if you want to go from a Python function to a running GPU job with minimal ceremony.
SkyPilot — shop 20+ clouds for the cheapest GPU
SkyPilot is a free, open-source (Apache 2.0) tool that automatically finds and runs your training jobs on whichever cloud or cluster has the cheapest available GPU capacity — it shops across 20+ clouds plus Kubernetes/Slurm clusters and fails over automatically if a provider runs dry. Since it runs inside your own cloud accounts, there's no markup. The catch: it isn't a GPU provider itself — you still need real cloud accounts with capacity behind it — and enterprise features like SSO and cost reporting require a paid managed layer on top. Pick SkyPilot if you already have accounts across multiple GPU clouds and want to stop manually price-shopping between them.
Production deployment: getting a model from laptop to live traffic
TrueFoundry — cloud-agnostic deployment and governance
TrueFoundry is a Kubernetes-based control panel for deploying, scaling and governing models and agents in production, without locking you into a single cloud. It has a free Developer tier (50K requests/month, 3 users), a Pro plan at $499/month (1M requests/month, 10 users), Pro Plus at $2,999/month, and custom Enterprise pricing for 10M+ requests. The pricing jump between Pro and Pro Plus for the same request ceiling is confusing — the real cost driver is seats and support level, so read the fine print before committing. Pick TrueFoundry if you're deploying across multiple clouds and want Kubernetes-level control without hand-rolling it yourself.
Vector storage: where your embeddings actually live
LanceDB — a database built for AI-scale multimodal data
LanceDB is an open-source (Apache 2.0) database purpose-built to store and search the images, text and embeddings AI systems train on and search through — it handles vector, full-text and hybrid search plus SQL filters natively, and is SOC 2/GDPR/HIPAA certified. The core is free and self-hostable; Cloud/Enterprise pricing isn't public. It's genuinely used in production at Netflix, Uber, Runway and Databricks, backed by a $30M Series A. The trade-off: it can be overkill if you just need simple vector search, and it requires learning the Lance data format to get full value. Pick LanceDB if you're building RAG or large-scale multimodal search, not just a basic vector lookup.
Eval: knowing whether a change made things better or worse
Parea AI — evaluation, feedback and monitoring in one
Parea AI lets teams building on top of LLMs test whether a prompt change actually improved answers, and watch what the model is doing once it's live. The free tier is genuinely usable for testing (2 members, 3,000 logs/month, 10 deployed prompts); Team is $150/month (3 members, 100K logs/month); Enterprise is custom with self-hosting. It's framework-agnostic with Python and TypeScript SDKs. The catch: $150/month is a big jump once you outgrow the free tier, and it's less natively integrated with LangChain than LangSmith is for teams already in that ecosystem. Pick Parea AI if you want eval, human feedback and production observability in a single tool rather than three.
Cost tracking: knowing what your AI usage actually costs
SuperPenguin — an itemized bill across every LLM provider
SuperPenguin shows exactly which customer, feature or team is burning through your OpenAI, Anthropic or AWS Bedrock budget, across 14+ providers in one dashboard, and it reconciles invoices to catch provider billing errors. Pricing: free up to $2K in managed spend, $30/month up to $5K, $200/month up to $20K (with forecasting), and custom Enterprise above that. It's Y Combinator-backed and claims sub-10ms SDK overhead. The catch: the free tier's $2K cap is really just a trial, there's no public GitHub repo to inspect the SDK, and pricing climbs quickly as spend grows. Pick SuperPenguin if nobody on your team can currently answer "which feature is driving our OpenAI bill."
Comparison at a glance
| Tool | Job | Starting price | Best for |
|---|---|---|---|
| Feast | Feature store | Free (OSS) | Teams with data engineering capacity |
| Hopsworks | Feature store + lakehouse | Free tier, then usage-based | Managed feature store + more |
| Modal | Serverless training/inference | Free ($30/mo credits) | Python-first GPU jobs |
| SkyPilot | Multi-cloud GPU scheduling | Free (OSS) | Price-shopping across GPU clouds |
| TrueFoundry | Production deployment | Free tier, then $499/mo | Multi-cloud deployment/governance |
| LanceDB | Vector/multimodal storage | Free (OSS) | RAG and large-scale search |
| Parea AI | Eval and monitoring | Free tier, then $150/mo | Eval + feedback + observability in one |
| SuperPenguin | AI cost tracking | Free (up to $2K spend) | Multi-provider LLM cost visibility |
None of these tools cover the whole stack alone — and that's normal. A realistic MLOps setup for a small team usually means one feature store OR a lakehouse, one training/compute layer, one deployment layer, and eval/cost tooling bolted on once things are actually in production and the bills start showing up. Start with whichever piece is causing the most pain today rather than building out the full stack up front.