# 🎉 What’s new in Elastic 9.3 **URL:** \ **Category:** Announcements\ **Created:** [February 3, 2026, 5:21pm UTC](https://discuss.elastic.co/t/what-s-new-in-elastic-9-3/384891 "2026-02-03T17:21:25Z")\ **Posts on this page:** 1\ **Page:** 1
**Author:** ![piotrprz](https://sea2.discourse-cdn.com/elastic/user_avatar/discuss.elastic.co/piotrprz/32/128141_2.png) [@piotrprz](https://discuss.elastic.co/u/piotrprz)\ **Post date:** [February 3, 2026, 5:21pm UTC](https://discuss.elastic.co/t/what-s-new-in-elastic-9-3/384891/1 "2026-02-03T17:21:25Z")
# 🎉 What’s new in Elastic 9.3 ![image](https://us1.discourse-cdn.com/elastic/original/3X/5/2/52d9c4da4231cddd1842270e7bebec28aa8655d7.png) Elastic 9.3 builds on the features introduced in 9.2 and focuses largely on **making them easier and more predictable to run in production** : better automation primitives, faster analytics, more efficient vector workloads, and measurable improvements in Elastic Cloud Serverless. Let's check some of the changes here. * * * ## From agentic concepts to operational automation Elastic 9.2 introduced the idea of agentic workflows: agents that can reason over Elasticsearch data using ES|QL and search. In 9.3, the focus shifts from introduction to **operational use**. - **Agent Builder** is now generally available. - **Elastic Workflows** (technical preview) adds a native orchestration layer that can be triggered by rules, alerts, or agents. The key change is that agents are no longer just interactive tools; they can now participate in **repeatable, automated flows**. Example: defining an ES|QL-backed tool an agent can use during a workflow: ```json POST kbn://api/agent_builder/tools { "id": "recent_errors", "type": "esql", "configuration": { "query": "FROM logs-* | WHERE log.level == \"error\" | SORT @timestamp DESC | LIMIT 10" } } ``` Example: invoking the agent via API and letting it select tools automatically: ```json POST kbn://api/agent_builder/agents/my-agent/chat { "messages": [ { "role": "user", "content": "Show me recent errors and summarize them" } ] } ``` * * * ## Search & AI: inference and relevance without running your own GPUs Elastic 9.3 expands the **Elastic Inference Service (EIS)** in ways that are mostly about operational convenience. 3 Jina models are now generally available through EIS: - [jina-embeddings-v3](https://jina.ai/models/jina-embeddings-v3) for multilingual embeddings - [jina-reranker-v2-base-multilingual](https://jina.ai/models/jina-reranker-v2-base-multilingual/) for semantic reranking - [jina-reranker-v3](https://jina.ai/models/jina-reranker-v3) for semantic reranking (new version) Example: creating an inference endpoint backed by Jina embeddings: ```json PUT _inference/text_embedding/jina_embeddings { "service": "elastic", "service_settings": { "model_id": "jina-embeddings-v3" } } ``` Indexing documents with embeddings generated by EIS: ```json PUT my-index { "mappings": { "properties": { "title": { "type": "text" }, "content": { "type": "semantic_text", "inference_id": "jina_embeddings" } } } } POST my-index/_doc { "title": "Elastic 9.3 overview", "content": "Elastic 9.3 improves automation, analytics, and vector search." } ``` * * * ## `bfloat16` vectors: reducing storage and memory pressure Elastic 9.3 adds support for storing dense vectors using `bfloat16` instead of 32-bit floating point values. This cuts vector storage roughly in half while preserving enough precision for many semantic search and retrieval workloads. In practice, `bfloat16` is most useful when: - vector dimensionality is high (e.g. 768, or more) - recall requirements tolerate small numerical error - memory pressure or disk footprint is a limiting factor Example: defining a dense vector field stored as bfloat16: ```json PUT bf16-vectors { "mappings": { "properties": { "embedding": { "type": "dense_vector", "dims": 768, "element_type": "bfloat16", "index": true } } } } ``` Indexing documents works the same way as with float vectors: ```json POST bf16-vectors/_doc { "id": "doc-1", "embedding": [0.0123, -0.9812, 0.4431, "..."] } ``` Querying `bfloat16` vectors does not require changes to the kNN query syntax; Elasticsearch handles the reduced precision transparently. `bfloat16` is especially effective when combined with disk-based vector indexing and on-disk rescoring, allowing large vector datasets to remain searchable without keeping all vectors resident in memory. * * * ## GPU-accelerated vector indexing (technical preview) Elastic 9.3 introduces **GPU-accelerated vector indexing** for self-managed deployments, built on NVIDIA GPUs and [cuVS](https://developer.nvidia.com/cuvs). This affects **indexing and maintenance** , not query execution. Reported improvements include: - up to **12× faster vector indexing** - up to **7× faster force-merge operations** - reduced CPU pressure during heavy vector ingestion * * * ## ES|QL: faster metrics queries and better time-series behavior ES|QL continues to mature as the main analytical interface in Elasticsearch. In 9.3, most changes focus on **performance and stability** , especially for metrics workloads: - **Sliding-window aggregations** reduce jitter in dashboards - **Exponential histogram support** improves distribution analysis - Metrics queries see **up to 5× lower latency** (technical preview) Example: smoothing a request rate using a sliding window: ```esql TS metrics | WHERE TRANGE(1h) | STATS avg(rate(requests, 10m)) BY TBUCKET(1m), host ``` Example: defining an exponential histogram field for OpenTelemetry metrics: ```json PUT otel-metrics { "mappings": { "properties": { "latency": { "type": "exponential_histogram" } } } } ``` Querying percentiles from histogram data using ES|QL: ```esql FROM otel-metrics | STATS p95 = PERCENTILES(latency, 95) BY service.name ``` * * * ## Elastic Observability: reducing cost without losing detail Several 9.3 changes in Observability are about **doing less work per byte** : - **Pattern-based log compression** (`pattern_text`, GA) can reduce log message storage by up to **50%** - **Elastic Streams** (technical preview) continues to evolve, using agentic techniques to structure logs directly from message fields - **Amazon Bedrock AgentCore integration** (technical preview) adds visibility into agent-based AI applications * * * ## Elastic Security: automation over repetition Security improvements in 9.3 focus on **reducing manual work in the SOC** : - **Entity AI Summary** (GA) provides structured summaries of entity risk - **Automatic rule migration** now includes QRadar (technical preview) - **Automatic gap filling** (GA) helps backfill missed detections - Deeper integration with Agent Builder and Elastic Workflows enables more automated investigation paths * * * ## Elastic Cloud Serverless: quieter changes, measurable effects Elastic Cloud Serverless doesn’t introduce new concepts in 9.3, but it does deliver **measurable improvements**. - Serverless is now available in **18 regions** - A backend infrastructure upgrade for AWS-backed projects results in: - up to **35% lower search latency** - **26% higher ingest throughput** These improvements are applied automatically and don't require configuration changes. For existing Serverless users, this is mostly a "things got faster" release rather than a behavioral change. * * * Elastic 9.3 is available now on [Elastic Cloud](https://cloud.elastic.co/) and for self-managed deployments. If you’re upgrading from 9.2, most changes are incremental but practical: faster queries, cheaper storage, fewer manual steps, and better automation hooks. As always, consult the release notes for full details and breaking changes.