turbopuffer

An i8 vector type landed in June, cutting storage and query cost by 75% compared to f32 for quantization-aware models.

Smaller vectors, lower cost — The new i8 type targets models already trained for quantized inference, making high-throughput retrieval meaningfully cheaper without a quality trade-off.

Namespace branching and fuzzy search — May brought instant copy-on-write namespace cloning via namespace branching, alongside a Fuzzy filter for typo-tolerant string matching. A C# API client shipped alongside an updated word_v4 tokenizer that runs roughly 3× faster than its predecessor, and API keys can now be scoped to list-namespaces only.

Sparse vectors, cross-cloud copies, and faster writes — April added sparse vector search support and cross-provider namespace copies between GCP and AWS. AWS write latency dropped 2.5× through a faster commit cadence. Full-text query length limits rose to 8,192 characters, attributes per namespace expanded to 1,024, and three new regions opened in São Paulo, South Carolina, and Belgium. Namespace pinning arrived as a way to hold frequently queried data in cache at high QPS without overpaying for idle time.