We went looking for the usual things here: integrations, pricing, which projects still ship. The vulnerability data is what stood out, and it is not close.
vLLM has 70 known advisories, 45 disclosed in the last twelve months. MLflow has 82, with 29 in the last twelve months. Both CRITICAL at maximum severity. The busiest message broker we track discloses 14 a year.
Search interest in "best mlops" is dominated by courses, books and certification. This page is not that. It is for people choosing what to put in production.
At a glance
| Tool | Job | Best for | Licence | Advisories (12m) |
|---|---|---|---|---|
| Ollama | Serving | Local models, fast start | MIT | 0 |
| vLLM | Serving | Throughput at scale | Apache-2.0 | 45 |
| LangChain | Orchestration | Breadth, ecosystem | MIT | 2 |
| LlamaIndex | Orchestration | Retrieval-heavy work | MIT | 1 |
| Pydantic AI | Orchestration | Typed, recent, clean | MIT | 0 |
| MLflow | Tracking | The incumbent | Apache-2.0 | 29 |
| Metaflow | Pipelines | Clean pipelines, no drama | Apache-2.0 | 0 |
| Langfuse | Observability | Open-source tracing | MIT | 0 |
| Portkey | Observability | Managed gateway | Proprietary | none published |
The advisory table
| Tool | Known | Last 12 months | Max severity |
|---|---|---|---|
| MLflow | 82 | 29 | CRITICAL |
| vLLM | 70 | 45 | CRITICAL |
| LangChain | 21 | 2 | CRITICAL |
| BentoML | 16 | 8 | CRITICAL |
| ZenML | 14 | 1 | CRITICAL |
| Ray | 12 | 6 | CRITICAL |
| LlamaIndex | 9 | 1 | CRITICAL |
| Haystack | 1 | 0 | CRITICAL |
| DSPy | 1 | 1 | MODERATE |
| Ollama, Langfuse, CrewAI, AutoGen, Pydantic AI, Metaflow, DVC, Kubeflow | 0 | 0 | none recorded |
A high count is not automatically a mark against a project. It reflects scrutiny and a disciplined disclosure process. What matters is the recent rate, because that is your patching commitment. This tooling is young, fast-moving, and routinely handles untrusted input: model files, serialised artefacts, remote weights. Deserialisation and remote-code-execution bugs are common in exactly that shape of software.
Serving
Ollama
MIT · 180,817 stars · shipped 2026-09-05 · 0 advisories · 70 integrations · from $20/month · score 73.2
Verdictthe default for local models and small deployments, and it deserves to be.
vLLM
Apache-2.0 · 91,636 stars · shipped 2026-09-09 · 70 advisories, 45 recent (CRITICAL) · 32 integrations · score 64.7
The throughput answer for serving at scale, and the highest recent advisory rate on this page.
Verdictstill the performance choice. Run it only with a real process for tracking and applying updates.
vLLM alternatives
vllm alternatives is one of our clicked queries, so by reason for leaving:
- The advisory rate worries you then Ollama (zero recorded) for smaller workloads, or a managed service so it is someone else's patching problem.
- You want it managed then Together AI, Baseten or Replicate.
- You want simpler local serving then Ollama.
Nothing open-source matches vLLM on throughput, which is the uncomfortable part of this answer.
Together AI, Baseten, Replicate
Baseten: usage-based, score 76.3, confidence 0.73. Together AI: from $0.00/tokens, score 75.7. Replicate: usage-based, 10 integrations, score 51.4.
Verdictthese remove the operational question entirely. Baseten is the strongest on our evidence.
Orchestration frameworks
LangChain
MIT · 145,850 stars · shipped 2026-09-04 · 21 advisories, 2 recent · from $39/seat/month · score 78.7
Verdictdominant on adoption, and a reasonable advisory profile for its footprint.
LlamaIndex
MIT · 52,129 stars · shipped 2026-08-19 · 9 advisories, 1 recent · from $50/month · score 69.7
Verdictthe retrieval-focused choice. mlflow vs langchain appearing in our data suggests people are not always clear which layer they are shopping for.
Pydantic AI, CrewAI, Haystack, AutoGen, DSPy
Pydantic AI: MIT, 19,902 stars, shipped 2026-09-12, 0 advisories. CrewAI: MIT, 58,372 stars, 0 advisories. Haystack: Apache-2.0, 26,478 stars, 25 integrations. AutoGen: CC-BY-4.0, 60,939 stars. DSPy: MIT, 37,899 stars.
VerdictPydantic AI and CrewAI both record zero advisories and shipped within the month, which is meaningful for a team nervous about the numbers above. AutoGen is licensed CC-BY-4.0, a content licence rather than a software licence, which is unusual enough to check with the project before building a product on it.
Tracking and pipelines
MLflow
Apache-2.0 · 27,843 stars · 82 advisories, 29 recent (CRITICAL) · 28 integrations · score 70.7
The incumbent, with the highest total advisory count here and the longest gap since shipping among actively-used tools.
Verdictwidely deployed and widely examined. Budget for patching.
Metaflow, DVC, Kubeflow, Ray, ZenML, BentoML
Metaflow: Apache-2.0, 10,261 stars, 0 advisories, shipped 2026-09-02. DVC: 15,867 stars, 0 advisories. Kubeflow: 15,856 stars, 0 advisories. Ray: 43,777 stars, 12 advisories, 6 recent. ZenML: 14 advisories, 1 recent. BentoML: 16 advisories, 8 recent.
VerdictMetaflow, DVC and Kubeflow all record zero advisories. BentoML's 8 recent disclosures put it in the same bracket as the serving tools above.
Observability
Portkey, Langfuse, Arize AI
Portkey: from $49/month, 16 integrations, score 78.9, confidence 0.76, the highest here. Langfuse: MIT, 34,480 stars, 0 advisories, shipped 2026-09-11. Arize: 33 integrations, the most.
VerdictPortkey scores highest overall; Langfuse is the open-source-first choice and the most recently shipped tool in the category.
Method, and what we could not verify
Confidence runs from 0.12 to 0.76, median around 0.57.
Dependency counts are unusable here and we have not used them. Ollama records 0 dependent repositories against 180,817 stars, vLLM 5 against 91,636, CrewAI 0 against 58,372, AutoGen 0 against 60,939, DSPy 3 against 37,899, Langfuse 0 against 34,480.
MLflow's recorded latest release is a rolling `model-catalog/latest` tag rather than a version.
LangSmith (0.12), Helicone (0.22) and Braintrust (0.31) should be read as unmeasured.
Advisory counts reflect published disclosures only. A closed-source hosted service showing none means nobody publishes them, which systematically flatters commercial products above.
We take no vendor payment for placement.
Evaluate these against your own constraints
Twenty-seven tools across four jobs, with wildly different security and maintenance profiles. Tell the evaluation agent what you are building.
Sources
Advisory data from advisories.ecosyste.ms, adoption from GitHub, both retrieved 2026-09-11. Pricing independently observed; each product page links the vendor page it was observed from.