Open-Source Machine Learning Engineer

Weaviate review 2026: pricing, features, pros & cons. Weaviate is the open-source vector database built around native hybrid search - combining vector similarity with keyword search in a single query. Here's an honest look at the new 2026 cloud pricing, self-hosting tradeoffs, and how it compares to Pinecone. Quick verdict. Overall Rating 100K objects $45/mo min Best for: Teams building RAG applications that need hybrid (vector + keyword) search and want the option to self-host without vendor lock-in. Less ideal for teams wanting the absolute simplest managed setup - Pinecone still wins there. What is Weaviate? Weaviate is an open-source vector database designed for AI applications, most commonly retrieval-augmented generation (RAG) pipelines. Its defining feature is native hybrid search: a single query can combine dense vector similarity search with sparse BM25 keyword search, weighted by an alpha parameter, so teams don't need to bolt a separate keyword search system onto their vector store. Because the core database is open source, Weaviate can be self-hosted via Docker or Kubernetes at no licensing cost, or run as a fully managed service through Weaviate Cloud. It integrates as a pluggable module with OpenAI, Cohere, HuggingFace, and other embedding and reranking providers, and ships with actively maintained Python, JavaScript/TypeScript, Go, and Java client libraries. In October 2025, Weaviate restructured its cloud pricing model entirely, replacing the old Serverless tier with a new Free / Flex / Premium structure billed transparently on vector dimensions, object storage, and backup storage. Weaviate pros & cons. Pros. * - Open source at its core: Weaviate can be self-hosted for free with the full feature set, giving teams a real exit from vendor lock-in that fully managed competitors like Pinecone don't offer * - Hybrid search built in: Weaviate natively combines vector similarity search with BM25 keyword search in a single query, weighted by an alpha parameter - useful for RAG systems where pure semantic search misses exact keyword or acronym matches * - Generous free cloud tier: The Free plan includes 100,000 objects, 1GB memory, and 10GB disk on an always-free cluster per user, enough to build and test a real RAG prototype before committing to a paid tier * - Transparent, dimension-based billing on paid tiers: Flex pricing is calculated from vector dimensions, object storage, and backup storage rather than opaque per-query pricing, making cost more predictable to forecast as usage scales * - Modular vectorizer and reranker modules: Weaviate integrates directly with OpenAI, Cohere, HuggingFace, and other embedding/reranking providers as pluggable modules, so you don't need to manage embedding generation in a separate service * - GraphQL and REST APIs with strong client libraries: Python, JavaScript/TypeScript, Go, and Java clients are actively maintained, and the schema/class model maps cleanly onto most application data models Cons. * - 2025 pricing restructure raised the entry point: Weaviate Cloud's paid tier starting price moved from $25/month under the old Serverless plan to $45/month minimum under the new Flex tier - a real cost increase for small production workloads migrating off the free tier * - Premium tier is a steep jump: after Flex ($45/month minimum), the next tier up (Premium, with 99.95% uptime and dedicated deployment options) starts at $400/month on a prepaid contract - there's a wide gap between hobby-scale and enterprise-scale pricing * - Self-hosting requires real operational expertise: running Weaviate yourself in Docker or Kubernetes for production means you own sharding, backups, and upgrades - the free-to-self-host framing understates the DevOps investment needed at scale * - Class/schema model has a learning curve: developers coming from a simple key-value or pure-vector mental model need to learn Weaviate's schema, cross-references, and module configuration before getting hybrid search and filtering working well * - Smaller managed-service footprint than Pinecone: Pinecone's fully managed simplicity and broader out-of-the-box framework integrations still win for teams that want zero infrastructure decisions and the fastest path from prototype to production Weaviate pricing 2026. Free. * - 100,000 objects * - 1GB memory / 10GB disk * - 1 collection * - 1 always-free cluster per user * - Best-effort availability Prototyping and small RAG projects before committing to a paid tier Flex. From $45/mo * - Unlimited objects * - Up to 1,000 collections * - Shared cluster with replication * - RBAC security * - 99.5% uptime SLA * - Pay-as-you-go dimension-based billing Production workloads that need real scale without a long-term contract Premium. From $400/mo * - Unlimited objects and collections * - Shared or dedicated deployment * - AWS, GCP & Azure coverage * - 99.95% uptime SLA * - Phone/Slack support + Technical Account Team Enterprise workloads needing dedicated infrastructure and faster incident response A Dedicated Enterprise tier is also available from $400/month prepaid with 99.95% uptime, 1-hour Severity 1 response, and HIPAA compliance on AWS. Weaviate can also be self-hosted for free. Weaviate vs Pinecone. | Feature | Weaviate | Pinecone | | Open source / self-hostable | | Fully open source | | Managed only, closed source | | Hybrid search (vector + keyword) | | Native BM25 + vector | | Sparse-dense hybrid supported | | Free tier | | 100K objects, always free | | Serverless free tier | | Entry paid price | $45/mo (Flex) | Usage-based, no fixed minimum on Serverless | | Enterprise tier | $400/mo+ (Premium/Dedicated) | Custom Enterprise pricing | | Setup complexity | | Schema/class model, more config | | Minimal config, fastest to production | | Vendor lock-in risk | | Low - can self-host anytime | | Managed-only, no self-host option | Who should use Weaviate? RAG apps needing hybrid search. Native vector + BM25 keyword search in one query makes Weaviate a strong fit for RAG systems where pure semantic search misses exact terms, product codes, or acronyms. Teams avoiding vendor lock-in. Because the core database is open source, teams can self-host on their own infrastructure at any point, giving an exit path that fully managed-only competitors don't provide. Cost-conscious startups building on the free tier. The always-free 100,000-object tier is real enough to build and test a production-shaped RAG prototype before committing to Flex or Premium. Enterprises needing dedicated deployment. Premium and Dedicated Enterprise tiers offer HIPAA compliance, choice of cloud provider, and 1-hour incident response for regulated or mission-critical workloads. Frequently asked questions. Is Weaviate worth it in 2026? For teams that want hybrid search (vector + keyword) built in and value the option to self-host without vendor lock-in, Weaviate is a strong pick - the always-free 100K-object tier is genuinely usable for prototyping. Teams that want the absolute fastest path to production with zero infrastructure decisions may still prefer Pinecone's simpler managed model, especially since Weaviate's October 2025 pricing restructure raised the paid entry point from $25/month to a $45/month minimum. How is Weaviate priced? Weaviate Cloud has three tiers: Free ($0, 100,000 objects, 1 always-free cluster), Flex (from $45/month minimum, pay-as-you-go based on vector dimensions, object storage, and backup storage, 99.5% SLA), and Premium (from $400/month prepaid, 99.95% SLA, shared or dedicated deployment). Weaviate can also be fully self-hosted for free via Docker or Kubernetes if you're willing to manage the infrastructure yourself. What is hybrid search in Weaviate? Hybrid search combines dense vector similarity search with sparse keyword search (BM25) in a single query, weighted by an alpha parameter you control. This matters for RAG systems because pure semantic search sometimes misses exact keyword matches, acronyms, or product codes that keyword search catches - hybrid search gets the best of both without running two separate systems. How does Weaviate compare to Pinecone? Weaviate is open source and self-hostable, giving teams an exit from vendor lock-in that Pinecone (managed-only, closed source) doesn't offer. Weaviate also has native hybrid search built into its core query model. Pinecone counters with a simpler, faster path to production and a broader out-of-the-box integration ecosystem across RAG frameworks. Cost-conscious teams comfortable with more setup often pick Weaviate; teams that want zero infrastructure decisions often pick Pinecone. Can I self-host Weaviate for free? Yes - Weaviate's core database is open source and can be self-hosted via Docker or Kubernetes at no licensing cost. The tradeoff is that you take on sharding, backup, and upgrade operations yourself, which requires real DevOps capacity at production scale. Weaviate Cloud's managed tiers exist specifically to remove that operational burden for a monthly fee. What changed in Weaviate's 2026 pricing? In October 2025, Weaviate restructured its cloud pricing entirely: old tier names were retired, the starting price for paid usage moved from $25/month (the old Serverless tier) to a $45/month minimum under the new Flex tier, and billing switched to a transparent model based on vector dimensions, object storage, and backup storage rather than opaque per-query pricing. Compare vector databases. See how Weaviate stacks up against Pinecone and other AI infrastructure tools.

DiScoFormer: One transformer for density and score, across distributions. Hugging Face introduces DiscoFormer, a transformer designed for density and scoring across distributions, signaling a new efficiency frontier in model scaling. Overview. DiscoFormer represents an architectural shift to unify density control and scoring objectives within a single transformer framework. The core idea is to optimize for both representation efficiency and predictive quality, potentially enabling larger runtime deployment on constrained hardware without sacrificing accuracy. The article also situates DiscoFormer within a broader trend of flexible model architectures aimed at reducing footprint while preserving generalization across distributions. Strategic implications: For developers and model builders, DiscoFormer may lower deployment barriers in edge devices and latency-sensitive environments. For platform providers, it creates opportunities to offer lighter, more portable models that still deliver robust performance in real-world tasks such as monitoring, decision support, and safety-critical inference. Industry context: The DiscoFormer narrative fits alongside ongoing explorations in model efficiency, quantization, sparsity, and density-aware training that have become central to practical AI scaling. It reinforces the thesis that architectural innovations may yield more impactful gains than scale alone in many use cases. Risks and considerations: New architectures require rigorous benchmarking across tasks, data regimes, and hardware backends. Adoption curves depend on tooling maturity and ecosystem support, including training-time tooling, quantization pipelines, and inference accelerators.

Google, Microsoft back draft AI agent discovery spec. Eleven corporations, together with Google, Microsoft, GitHub, and Hugging Face, have printed Agentic Useful resource Discovery (ARD). The open specification units out how AI brokers discover and confirm instruments, expertise, and different brokers throughout the net. The contributors launched the draft spec on June 17, together with reference implementations from a number of of them. ARD is licensed beneath Apache 2.0 and builds on the AI Catalog information mannequin maintained by a working group beneath the Linux Basis. The total record of contributors additionally contains Cisco, Databricks, GoDaddy, NVIDIA, Salesforce, ServiceNow, and Snowflake. What does ARD resolve? The spec units out to resolve a coordination drawback. In the present day, an agent needs to be wired to every instrument, MCP server, or API it makes use of forward of time. As extra corporations publish their very own capabilities, that pre-wiring stops scaling. ARD strikes discovery right into a search step that occurs at runtime. The shift primarily impacts corporations that publish instruments and brokers, not typical content material websites for now. How ARD works. ARD depends on two items, which the spec calls catalogs and registries. A corporation publishes a catalog, an ai-catalog.json file hosted at a well known path by itself area, that lists the instruments, MCP servers, brokers, or APIs it makes accessible. Registries then crawl these catalogs, index them, and reply discovery requests from brokers in plain language. As a result of every catalog sits by itself writer's area, the spec makes use of area possession to confirm who printed it. For manufacturing use, publishers can connect belief metadata so an agent or registry can affirm the writer's cryptographic id earlier than connecting. As soon as a functionality is chosen, ARD palms off and the agent connects straight utilizing the instrument's personal protocol. Similar-Day implementations. A number of contributors shipped working instruments constructed on the spec the identical day. GitHub launched agent finder, which lets Copilot uncover matching MCP servers, expertise, instruments, and brokers from a selected registry, with customers controlling what will get linked. Hugging Face launched a Uncover Instrument that searches expertise and MCP servers throughout ARD companies. Cisco tied the spec to its AGNTCY Agent Listing, an open supply venture beneath the Linux Basis. The discharge continues a run of open specs aimed on the machine-readable layer of the net. Google printed the Open Knowledge Format, a spec for sharing organizational data between AI programs, two days earlier. The sample throughout these efforts is similar. Every asks you to publish a structured file beneath your personal area so AI programs can use what you expose, with out anybody wiring the connection by hand. The place Google suits. Google's position facilities on Agent Registry, a part of its Gemini Enterprise Agent Platform. The corporate stated Agent Registry will host and search agentic sources and deal with enterprise governance. Native ARD help within the platform is deliberate for the approaching months, which Google stated would let organizations join inside registries to the broader community. That help is just not reside but, and ARD is a specification moderately than a Google Search characteristic. Why this issues. The cut up will depend on what you publish. ARD is for publishers of callable capabilities, the APIs, MCP servers, and brokers that software program connects to. An organization that publishes instruments has a transparent methodology for being discovered and trusted by brokers. A typical content material website has no clear motion to take right this moment. The worth of this effort is debated. Google's John Mueller has argued that LLM programs can't use recordsdata like llms.txt to differentiate one website from one other, and suggested specializing in present wants moderately than future agent-oriented methods. ARD targets instruments and brokers, not content material, elevating questions on constructing now for programs that will or might not generate site visitors later. Trying forward. The spec is a v0.9 draft, and the contributors are inviting modifications by way of the venture's GitHub repository. Its attain will depend on registries that may crawl and index catalogs at scale, and that ecosystem continues to be in its early levels. Google's Agent Registry help is months out. If that community develops, the benefit primarily goes to corporations providing instruments and brokers that others want. The early agentic-web features from Google trace at this. The rapid concern is whether or not your present platforms and instruments will undertake ARD and what they may require you to publish.