How Much Are Free Qdrant Credits Worth?
Qdrant offers $2,000 in free credits for startups, applied against Qdrant Cloud, the managed version of its open source vector database.
That number is modest next to the six-figure grants elsewhere in this category, and it is worth reading correctly rather than dismissing. Qdrant is Apache 2.0 licensed and self-hostable, so the paid product is managed hosting, not the engine. The credit is a hosting budget, not an access key.
Eligibility depends on stage and funding, and infrastructure program terms move more than model program terms do. AI Perks tracks the current version alongside $7.7M in credits across 194 companies.
What Qdrant Actually Does
Qdrant stores embeddings alongside structured metadata and returns the nearest matches to a query vector in milliseconds. It is written in Rust, ships as a single binary, and behaves the same on a laptop as in production.
Every product that answers questions over private documents, searches by meaning rather than keyword, recommends similar items, deduplicates records, or gives an agent long term memory needs this layer. You either run it or you rent it.
Three properties separate Qdrant from the rest of the category:
Payload is a first class citizen. Metadata lives with the vector and is indexable, so filters are part of the query rather than a second pass over the results. More on why that matters below.
It is one binary. Local development parity is a real operational advantage. You can run the identical engine in a Docker container on your machine and in a managed cluster, which removes a category of "works locally" surprises.
The license is permissive. Apache 2.0 means self-hosting stays available permanently. That is negotiating leverage rather than a plan: knowing you could move keeps the managed bill honest.
Filtered Search Is Where Retrieval Products Break
Almost no production query is "find similar". It is "find similar, inside this customer's documents, from the last 90 days, of this type". That combination, not raw similarity, is where most vector search quietly fails.
The failure has a specific shape. Naive implementations post-filter: retrieve the top 100 nearest vectors, then discard the ones that do not match the filter. If the filter is selective, you get three results back, or zero, and the product looks broken even though the index is fine.
The alternative, pre-filtering by brute force over the matching subset, is correct but slow once the subset is large. Qdrant's answer is to apply filter conditions during graph traversal and to keep the search graph connected under filtering, backed by payload indexes on the fields you filter on. The engineering detail matters less than the consequence: filtered queries stay both fast and complete.
This is the single most important property for anyone selling B2B software, because every query in a multi-tenant product is filtered by tenant. If your retrieval degrades under filters, it degrades for every customer at once.
How Vector Database Cost Behaves at Scale
Vector search cost is driven by how many vectors you store multiplied by their dimensions, not by how many queries you run. An idle index over a large corpus is expensive. A busy index over a small corpus is cheap.
That is the inverse of how model APIs bill you, which is why founders misforecast this line item so consistently. Here is the arithmetic for one million vectors, before index overhead:
| Storage mode | Bits per dimension | 1M vectors at 768 dims | 1M vectors at 1,536 dims |
|---|---|---|---|
| float32 in memory | 32 | ~3.1 GB | ~6.1 GB |
| Scalar quantization (int8) | 8 | ~0.8 GB | ~1.5 GB |
| Binary quantization | 1 | ~0.1 GB | ~0.2 GB |
| Memory-mapped on disk | 32, paged | RAM holds the index | RAM holds the index |
Treat those as a floor. Graph overhead and payload sit on top.
Two levers change the picture more than anything you will do to your code. Quantization compresses stored vectors, and Qdrant can rescore the shortlist against full precision vectors to recover most of the lost recall, which makes it far less of a downgrade than it sounds. Dimension choice is the other: many embedding models now support shortened outputs, and halving dimensions halves the whole table.
The practical takeaway is that chunking strategy is a cost decision, not just a quality decision. Splitting documents into smaller pieces multiplies your vector count and your memory bill with it. AI Perks tracks credits across this layer so you can test at real scale before the invoice teaches you.
What a $2,000 Credit Is Actually For
A $2,000 grant is not a year of runway. It is enough to run your real corpus at real scale long enough to learn what production will cost, which is worth more than the money.
Teams that size a vector bill from a 10,000 row sample are usually wrong by an order of magnitude. Recall, filter performance, and the quality hit from quantization all change with corpus size, and none of them extrapolate cleanly from a toy index.
So spend the credit as a measurement, not a discount:
Load everything, not a sample. The whole point is to see behaviour you cannot see at small scale.
Query with filters on. Test the query shape your product will actually issue, tenant filter included.
Measure recall at your quantization setting. Decide the memory tradeoff with a number instead of a preference.
Then price self-hosting against managed. With a real index size, that becomes arithmetic rather than an argument.
Timing matters as much as the amount. Credits are time-boxed, so activating a grant while you are still choosing a chunk size burns most of it on an empty index. Line the clock up with the moment you have both a corpus and traffic. The current programs and their terms sit at AI Perks.
What Qdrant Credits Stack With
Retrieval bills in three places: creating the embeddings, storing and searching them, and generating the answer. A vector database grant covers one of the three.
| Layer | How it bills | Where credits exist |
|---|---|---|
| Embeddings | Once per document, again on every model change | OpenAI, Cohere, Voyage, Mistral |
| Vector storage and search | Continuously, driven by memory | Qdrant, Pinecone, Weaviate |
| Generation | Per answer, per user | Anthropic, OpenAI, Google, xAI |
| Compute and hosting | Continuously, driven by uptime | AWS, Google Cloud, Azure, Modal |
The embedding row is the one that catches people. You embed the corpus once to index it, then re-embed all of it every time you change model. An embedding upgrade is a full reindex, and on a large corpus that is a real invoice rather than a background job.
Holding credits across several rows is what lets a team change its mind without a budget meeting. That is the actual argument for stacking.
What Founders Get Wrong About Vector Databases
The two most common mistakes are adopting one before the product needs it, and treating retrieval quality as a model problem when it is usually a filtering or chunking problem.
Five failures worth avoiding:
Starting too early. Postgres with pgvector handles a few hundred thousand vectors and straightforward queries, and it removes a service from your stack. A dedicated vector database earns its place at scale, at sustained query volume, or when filtered search starts returning too little.
Ignoring filters until multi-tenancy arrives. By then the query patterns are baked in and the fix is a migration.
Treating quantization as a downgrade. It is the difference between a 6 GB index and a 1.5 GB one, and with rescoring the recall cost is often negligible. Measure it rather than assuming.
Confusing open source with free. Self-hosting removes the license fee, not the memory bill or the on-call rotation.
Forgetting the embedding side. Storage credits cover half a retrieval bill. Founders who search for both at once fund a year. The rest fund a quarter. Both sides are catalogued at AI Perks.
Frequently Asked Questions
How much are Qdrant credits worth?
Qdrant offers $2,000 in free credits for startups, applied to Qdrant Cloud rather than to the open source engine, which is free anyway. How far it goes depends on corpus size and vector dimensions rather than query volume, since cost is memory-driven. Current terms and eligibility are tracked at getaiperks.com.
Is Qdrant free to use?
Yes, in the sense that matters: Qdrant is Apache 2.0 licensed and you can run it on your own hardware with no license fee. What costs money is Qdrant Cloud, the managed service, plus the memory your index consumes wherever it lives. Self-hosting trades money for operational work.
Do I need a vector database, or is pgvector enough?
For a few hundred thousand vectors and simple queries, Postgres with pgvector is usually enough and removes a service from your stack. A dedicated vector database earns its place at tens of millions of vectors, sustained query load, or heavy filtered and multi-tenant search, where naive implementations degrade sharply.
How much RAM does Qdrant need?
More than most teams budget for. Uncompressed 32-bit vectors need roughly 4 bytes per dimension per vector, so a million 1,536-dimension vectors is about 6 GB before index overhead. Scalar quantization cuts that close to fourfold and binary quantization much further, which makes quantization a planning decision rather than an optimisation.
Can I combine Qdrant credits with OpenAI or Anthropic credits?
Yes, and they cover different bills. Qdrant covers storing and searching your vectors, embedding providers cover creating them, and model providers cover generating the answer. Teams holding credits across all three cover a realistic first year of a retrieval product. See which combinations exist at getaiperks.com.
What are the main Qdrant alternatives?
Pinecone, Weaviate, Milvus, Chroma, and Postgres with pgvector. Qdrant's distinguishing points are its Rust engine, its filtering model, and a permissive license that keeps self-hosting available as a genuine exit rather than a bluff. AI Perks tracks credit programs across the whole category at getaiperks.com.
Store the vectors. Let someone else pay for the memory.