Qdrant and Minima Test a Faster Agentic RAG Stack
A controlled vendor benchmark combines hybrid retrieval with compressed Qwen inference, raising successful task throughput while exposing important workload and cost boundaries.
Qdrant and Minima published a controlled agentic retrieval-augmented generation benchmark on August 13, 2026, combining Qdrant hybrid retrieval and reranking with Minima's compressed Qwen3.6-27B serving stack. Their joint configuration completed 3,158 successful tasks per GPU-hour against 1,081 for the dense-retrieval, BF16 baseline, a reported 2.92-fold increase on the tested workload.
What the benchmark changed
The Qdrant and Minima engineering report separates two sources of wasted work inside an agent loop. The retrieval layer can return weak evidence and force another search. The inference layer can consume more memory and time than the task requires. The test changes both, which is why its headline result cannot be credited to either component alone.
On the retrieval side, the baseline used dense search and passed the top 16 results into the agent. The joint stack added BM25 sparse search, reciprocal-rank fusion and ColBERT-style late-interaction reranking, then passed the top eight results. Payload filters remained in the query path. The first retrieval was judged sufficient for 87% of tasks in the joint setup, up from 72%, while mean retrieved context fell from about 5,200 to 2,300 tokens.
On the inference side, Minima served Qwen3.6-27B with compressed weights and tiered key-value cache formats on a single 96 GB NVIDIA RTX PRO 6000 Blackwell GPU. The report says the baseline and compressed configurations used the same model checkpoint, hardware, sampling settings, prompt, concurrency and endpoint. Qdrant ran on a separate host, and query encoding ran on a separate CPU-only service.
What was measured
The evaluation covered 1,800 multi-step tasks from SciFact, FiQA and HotpotQA plus a fourth payload-filtering set created by Minima. It replayed 10,000 complete agent episodes and ran a separate 50,000-query adversarial filtering test for each retrieval condition. Each agent episode allowed no more than two Qdrant calls, and final answers were capped at 256 output tokens.
Under those conditions, raw throughput rose from 1,350 to 3,750 tasks per GPU-hour. Median task latency fell from 21.3 seconds to 7.7 seconds, while reported grounded-task success moved from 80.1% to 84.2%. Applying those success rates produced the headline move from 1,081 to 3,158 successful tasks per GPU-hour.
The authors also estimate a lower GPU cost per successful task. That accounting uses an assumed hourly GPU rate and explicitly excludes the Qdrant host and embedding services. It is therefore a comparison of the dedicated inference GPU in this setup, not a complete bill for a production RAG system.
Why the design is useful
The study's practical value is its attempt to measure the whole agent episode rather than quote vector-query latency or model throughput in isolation. A slower but more selective retrieval step can still reduce total work when it prevents a second search and shrinks the context sent to the model. Compression can then increase the number of concurrent or completed generations the same inference GPU can handle.
That systems view is relevant to teams optimizing agentic RAG. It suggests measuring evidence sufficiency, repeated retrieval, context volume, tool validity and completed-task throughput together. Optimizing a single component may move cost elsewhere or leave end-to-end latency unchanged.
Limits before generalizing the result
This is a vendor-authored report from the companies whose products are being evaluated, not an independent replication. One of the four task sets is internal, the agent has a fixed two-search limit, answers are short, and the hardware and model are specific. The study does not show the same multiplier on private enterprise corpora, longer answers, different rerankers, other GPUs, another model family or workloads with more complex tool use.
The configurations also change retrieval strategy and inference representation together. The report provides intermediate comparisons, but a buyer still needs workload-level tests to determine which improvement transfers and whether the extra retrieval services, memory policies and operational complexity are worthwhile.
Status
Learning. The published engineering report provides detailed methods and bounded numerical results for a reproducible comparison, but the evidence is first-party and jointly authored by Qdrant and Minima. Internal confidence is medium pending independent reproduction on additional corpora, models and hardware.
Sources
Update note: Last reviewed 2026-08-17. We will revise this post if independent teams reproduce the result or the authors publish code, task data and broader hardware comparisons.
Sources
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.