Saleor News

OpenAI's first model rated Critical for cybersecurity under the Preparedness Framework: it can find previously unknown vulnerabilities and exploits without step-by-step human guidance. Context: 1.05M tokens; output: 128K; cutoff: 30.04.2026; price: $10/$50 per million tokens. Limited preview for trusted partners, with closed weights.

The third Flash release in 6 weeks (after 3.6 on 21.07 and 3.7 on 13.08): the same speed and price as 3.7, but improved reasoning and coding. The Cyber version is a separate model for defenders available through the Fairwind access program; it found a 13-year-old bug in Chrome. Cloud only, with closed weights.

New SEARCH_ACTION mount point puts app actions in Dashboard command palette; widgets expand into full POPUP extensions via OpenPopup; new RedirectToApp action lets apps hand off to each other by manifest identifier instead of hardcoded URLs. Requires Dashboard 3.23.28 + app-sdk 1.15.0.

Previously, running AI Assistant self-hosted required manually enabling the instance-ai module and configuring the sandbox, model, and web search through environment variables. In 2.35, the module is enabled by default, installation takes one command, and the model provider (Anthropic, OpenAI, OpenRouter, or any OpenAI-compatible endpoint, including local ones), sandbox, and web search are configured directly in the UI.

Customers now segmentable and carry structured data via same attribute system already used for products/models; new MANAGE_CUSTOMER_TYPES_AND_ATTRIBUTES permission gates it.

The March roadmap has been expanded from 4 to 5 priorities. Agent identity and SDK developer experience have been added as a separate item; messaging primitives (Tasks, subscriptions/listen, progress) are being extended to server-initiated events via webhooks/channels instead of polling, and the local stdio transport is proposed to be unified with remote HTTP. The release follows the 2026-07-28 specification, the largest revision since the protocol launched.

A production RAG infrastructure example: PostgreSQL full-text plus semantic search with Qwen3-Embedding-0.6B (256-dimensional Matryoshka vectors), corpus embedding in batches on HF Jobs (L4, about 75 articles/sec), Storage Buckets as a versioned layer between compute and the database, and merging via Reciprocal Rank Fusion with a lexical-only fallback on timeout.

Nine security advisories in one release: 6 High (including an Expression Sandbox Escape via a prototype leak in $fromAI leading to Host RCE — GHSA-9x83-43r8-5hwc; RCE in the Git node; credential leaks from Strapi, SeaTable, and Mailcheck nodes; SSRF and local file reads through Gmail/Brevo) and 3 Medium (including query injection in Elasticsearch/Firestore nodes). Fixed in v1.123.73, v2.35.4 (stable), and v2.36.2 (beta); Cloud was patched automatically, while self-hosted instances require manual updates.

The new MultiVectorEncoder class brings ColBERT-like late interaction on top of dense/sparse models, already in the library. It is more precise on nuances (token-to-token MaxSim instead of compressing into a single vector), but the index is orders of magnitude heavier: 311.5 MB vs 7.5 MB dense on 4,874 Natural Questions passages. 40+ open checkpoints (LateOn, mLateOn, ColBERT, ColPali) are available on HF Hub.

Automatic quota increases at no extra cost: the Free plan grows from 5 to 25 GB, Professional from 1 GB to 10 TB, and Enterprise up to 120 TB; covering files, call recordings, and task and deal attachments.

An application with two modes: Cowork performs work tasks (reports, planning, and data analysis) itself rather than just advising; Code uses the Vibe coding platform to build applications without programming. The BitrixGPT model is closed, runs only on the vendor's servers in CIS, and its weights are not available for self-hosting.

A production model for coding and agents: DeepSWE v1.1 scores 65.3% versus 49.0% for 3.6 Flash, with a context window of 1,048,576 tokens. Video analysis reduces token usage by up to 88% and cost by up to 66%. API/cloud only, with closed weights.

A single visibility and control layer for agents in Anypoint Platform: verifiable agent identity, isolation of provider credentials, emergency agent shutdown, and auditing. Runs on top of AWS, Azure, and F5.

FinOps tools for AI: request cost calculation at the actual rate, budget limits per agent, and expense breakdown by agent. Model Proxy is in GA; the remaining features (Cost Optimization, Model Wallet, Unified Dashboard) will follow later in the quarter.

C# adds labeled break/continue, improved union pattern matching, and exhaustiveness checking for constrained type parameters. NativeAOT CLI and the MSBuild server are enabled by default in the SDK. EF Core brings improvements to query translation and Cosmos DB.

In TAdviser’s MDM market review (about 900 projects, market size of 6-7 billion rubles by the end of 2025), most MDM deployments in CIS were attributed to Datareon solutions. This is TAdviser’s estimate, not an independent market share audit.

Routine maintenance releases: 3.14.7 includes about 499 bug fixes from 86 contributors, and 3.13.15 includes about 400 bug fixes. A regular support cycle with no announced critical security fixes.

Columnar Mode (tech preview) stores a field once in a column store without an inverted index, letting Elasticsearch operate as a fully columnar store for search and analytics with less disk usage. VectorDB index mode automatically calibrates vector quantization without manual index tuning.

The first release candidate of 3.15 is out, with the final release planned for 2026-10-01. Confirmed features include lazy imports (PEP 810), built-in frozendict (PEP 814), a Sentinel type (PEP 661), unpacking in comprehensions (PEP 798), and a new sampling profiler (PEP 799); JIT has added about 8-9% performance on x86-64 Linux.

Django dependency upgraded to 5.2.17, fixing CVE-2026-15307 (server-side file-write/request forgery via spatial lookups), CVE-2026-15337 + CVE-2026-15830 (DoS), CVE-2026-15920 (XSS via URLField in admin).

The cause was a misconfigured eval environment with partner Irregular: test machines were given real internet access despite a system prompt stating they were isolated. The models treated the CTF target as legitimate and attacked the infrastructure of three real companies; the incident was found while reviewing 141,006 test sessions.

Yandex announced a general-purpose AI agent built on Yandex AI Studio to handle 30–50% of routine work for managers, analysts, and accountants: integration with AmoCRM and Kontur.Focus, scheduled background mode, and no-code custom skills. Release is planned for late September 2026; pricing has not been disclosed.

Fixed a bug where string fields containing dates were incorrectly converted to DateTime when saved to the database. Critical integration issues with 1C and databases on Linux were also resolved, and authentication error messages were unified to resist credential stuffing.

The initialize/initialized handshake and session ID have been removed, allowing requests to reach any server instance behind a load balancer without shared storage. Multi Round-Trip Requests replace streaming server-to-client calls; header-based routing (Mcp-Method/Mcp-Name) enables gateways to route without parsing the JSON body; list results are cacheable (ttlMs/cacheScope); and authorization has been strengthened (RFC 9207 and Client ID Metadata Documents instead of Dynamic Client Registration). Roots, Sampling, Logging, and HTTP+SSE transport are deprecated with a 12-month window; all Tier 1 SDKs (TS, Python, Go, C#) have been updated, while Rust is in beta.

The largest revision since launch. Protocol-level sessions (Mcp-Session-Id), the initialize/notifications/initialized handshake, ping, and SSE resumability were removed; cross-call state is now carried in explicit handle arguments. server/discover was added, the MRTR pattern replaced server-initiated requests, and resultType became mandatory. Roots, Sampling, Logging, and the HTTP+SSE transport are marked Deprecated (at least 12 months before removal), and Dynamic Client Registration gives way to Client ID Metadata Documents.

The process handling service and message store have been moved to core cluster services with auto-deployment to all nodes; internal storage and indexing in DatareonDB have been redesigned. The 1C connector can now record delete events in information registers. Default passwords have been removed - password change is required on the first sign-in to the configuration center.

Built-in analytics app (rolling out via Saleor Cloud) surfaces revenue, orders, returns, fulfillment trends, low-stock alerts in Dashboard, with agent-ready data layer for AI agents to query store metrics via permission-aware read-only tools.

GigaChat 3 Ultra became available to individuals through the public API in a free Freemium mode (the limit was raised to 365 million tokens by 22.07.2026). Cloud access only; the announcement does not mention published weights for this specific build.

MuleSoft publishes Mule skills from the mulesoft-dx repository as the @salesforce/mulesoft-vibes-skills package. The same instructions work in MuleSoft Vibes, Claude Code, Cursor, Windsurf, Cline, and Codex. Skills are activated on demand: a skill loads only when a request matches it, saving context.

An update to a closed model, API-only. Muse, not Llama, is established as Meta's frontier line.

Three models differentiated by price and speed: Sol leads in coding, science, and cybersecurity (Agents' Last Exam 53.6, ahead of competitors); Terra matches GPT-5.5 at half the price; Luna is 80% cheaper than Sol. Ultra mode coordinates multiple agents in parallel. Cloud only, with closed weights.

A new agentic UI layer lets users manage a fleet of agents that enrich, validate, and orchestrate product data among themselves. The platform is positioned as moving from a system of record to an agentic system of execution.

A 432B MoE model (28B active) using a hybrid MLA + GatedDeltaNet architecture (linear attention): long-text generation is 4× faster, while the model is almost twice as compact as GigaChat 3.1 Ultra with comparable code and math quality. Weights are available on HuggingFace (ai-sage/GigaChat3.5-432B-A28B), but the model was not yet available in the official cloud API at developers.sber.ru on the release date.

A beta for Pro/Max/Team/Enterprise built on existing Claude models: more than 60 preconfigured scientific databases and connectors (genomics, single-cell, proteomics, structural biology, and cheminformatics) in one workspace.

The most agentic Sonnet model to date: it plans multistep tasks and works with browsers and terminals without constant human supervision. Since July 1, it has been the default model for Free/Pro, replacing Sonnet 4.6. Promotional pricing is $2/$10 per million tokens through 31.08.2026. Closed weights.

Declarative pipelines as contracts, pipeline generation through coding agents (Claude Code, GitHub Copilot), and Qlik MCP Server, which exposes Qlik Talend Cloud governance objects (data products, lineage, catalog) as MCP tools, are now generally available in Qlik Cloud / Qlik Talend Cloud.

orderGrantRefundCreate/Update mutations accept reasonReference per order line, stored on OrderGrantedRefundLine — single refund spanning several items can carry different reason per line.

New HOMEPAGE_WIDGETS mount point lets apps render widgets in configurable Dashboard home grid; new App Bridge WidgetResize action added. Separate Anonymizer app strips personal customer data for GDPR erasure requests and dev-DB sanitizing.

Python 3.15.0 beta 3 is out - the feature freeze has passed, and stabilization is underway before release.

Manage the platform without opening the Anypoint UI through a native Slack agent, Claude Code, Microsoft Teams, or any MCP-compatible tool (such as Cursor): view applications, deploy code, apply policies, and generate governance reports in natural language.

Studio and Runtime now support Java 21 (a direct upgrade is possible only from 8.0.1-R2024-05 or later). Apache Syncope is no longer included in the ESB installer and must be installed separately from the Apache Syncope project. The current combination is Syncope 4.1 + Java 21 + Tomcat 10; Syncope 2 is incompatible with TAC/STS starting with this release.

Each workflow run creates a root span with status, duration, and error, plus an optional child span for each node. Enabled with two env variables, exported to any OTLP collector (Jaeger, Tempo, Datadog, Honeycomb, Grafana Cloud) with no vendor lock-in.

Agent Harness automatically compacts chat history when the token limit is exceeded and includes built-in file, todo, and web-search providers. CodeAct (agent-framework-hyperlight) combines multiple tool calls into one program in an isolated micro-VM: 52.4% lower latency and 63.9% fewer tokens in test scenarios. Foundry Hosted Agents provide scale-to-zero deployment with file system persistence between sessions.

Alongside static analysis, expert review, and a public vulnerability registry, AI code audits now cover modules (>2,000 solutions in the marketplace) against CWE/OWASP Top 10/CVSS; vulnerable versions are blocked from publication until fixed.

Opus 4.8 was released 41 days after 4.7 — Anthropic’s fastest update cycle. Dynamic Workflows (research preview) lets the model build and modify a graph of hundreds of parallel subagents for a complex task. Weights are closed; access is available through the API, Bedrock, Vertex AI, and GitHub Copilot.

Yandex released Alice AI LLM Flash: according to the vendor, it outperforms GPT-5.4 mini in 56% of business cases and costs approximately 5 times less than the lineup's top model. It was trained entirely on Yandex data; weights are not published and access is available only through the Yandex Cloud API.

25 KIPs, 600+ commits since 4.2.0. KIP-1244 deprecates the streams-scala module (removal in 5.0). KIP-1237 deprecates the group.coordinator.rebalance.protocols configuration. KIP-1280 moves MirrorMaker to the unified KIP-877 metrics mechanism; the old MirrorMaker metrics are also deprecated.

n8n 2.22.0: the agent connects to some external MCP servers (Apify, Linear, Notion, and others from the official MCP registry) by selecting them from the node panel and logging in, without manually creating an MCP Client node and credential.

Verified, GS1-aligned product data (attributes, images, and dimensions) from Syndigo is connected directly to Blue Yonder planning and execution workflows—one product record update in Syndigo propagates across the entire Blue Yonder Network without manual data cleansing by procurement teams. Announced at the Blue Yonder ICON conference.

BitrixGPT 5.5 now supports long context, files/images, and search (the model itself is closed, with closed weights). OpenClaw can be connected separately as a ready-made external open-source AI agent for tasks/CRM within the user's access permissions (open weights/open source).

The largest release by number of fixes (501 in Open Source, 560 in Commerce); Laminas MVC was replaced with native PHP, Zend_Cache with Symfony Cache, TinyMCE with HugeRTE, and OAuth was migrated to native PHP functions. PHP 8.2 support has ended; PHP 8.4 or 8.5 is required.

The company's valuation more than doubled in less than a year (from $2.5 billion). Under a commercial agreement, n8n is being natively integrated into Joule Studio, SAP Business AI Platform's agent-building environment.

The PHP team released patches simultaneously for all four supported branches, including fixes for PHP-FPM XSS and SQL injection.

Node 26 includes the Temporal API (Temporal.PlainDate, Temporal.ZonedDateTime) without flags as a replacement for the outdated mutable Date; V8 is updated to 14.6 (Chromium 146), and Undici to 8.0. Current release, entering LTS in October 2026.

Connectors for finance, marketing, CX, ITSM, security, DevOps, IT, and media systems. A continuation of the connector release series after 50 connectors in winter 2026.

The AI client can now create and edit workflows directly in an instance through the built-in MCP server. The intermediate representation is TypeScript that passes type-checking and compilation before application, rather than raw JSON. Public preview; n8n 2.18.4+ is recommended.

An update to the flagship lineup after 5.1, serving as a price-to-quality benchmark for the later GPT-5.6 Terra (the same intelligence level at half the price). API/ChatGPT only, with closed weights.

Users can now connect their own OpenAI or Anthropic API key instead of Akeneo’s built-in AI model, providing control over provider selection for compliance and security. Extension Platform enables full-stack extensibility with backend hosting through Upsun.

Pimcore is included in the 2026 Gartner MQ for Master Data Management Solutions report as a Niche Player; analysts highlight its Data Spine strategy, which unifies structured and unstructured data.

Business users run Synapse agentic workflows for item creation, onboarding, enrichment, and syndication through natural-language chat instead of the PIM interface. Announced at Syndigo Connect.

Studio moved from beta to GA after two years of development and replaces the legacy ExtJS admin panel; version 2025.4 remains a transition bridge with both UIs running in parallel. The foundation for the Pimcore Agent SDK and Agentic PXM model has been laid, with agents performing tasks within human-defined boundaries.

At Inspire 2026, Data Spine was introduced as a unified governance layer connecting PIM/MDM, content platforms, and AI: every human or agent action is checked against business rules, workflows, and access rights.

Routes in Talend Studio can call tools from any MCP-compatible server, including Qlik MCP Server; the context window has been expanded for complex agentic scenarios. Agentic Routes have existed since early 2025, so this is an enhancement, not a new feature.

The FNS Commission (Decision No. 936-039 dated April 3, 2026) confirmed Taxcom’s compliance with requirements for EDI operators and extended its status until April 3, 2028; the service supports roaming with 33 operators.

A unified API for encoding and comparing text, images, audio, and video with a single model; rerankers have also become multimodal (Qwen3-VL, BAAI BGE-VL, NVIDIA Nemotron—all with open weights). The retrieve-and-rerank pattern now works with mixed content, enabling RAG for visual documents and screenshots, not just text.

The first proprietary MSL model: API-only, with no published weights. A break from Llama's distribution model of downloadable open weights and self-hosting.

AutoGen and Semantic Kernel merged into a unified SDK with stable APIs and long-term support. An MIT license, open source, and open weights are not required — the framework orchestrates any models: Foundry, Azure OpenAI, OpenAI, Anthropic Claude, Bedrock, Gemini, and Ollama. MCP and A2A support enables cross-platform agent interaction.

Sber released GigaChat-3.1-Ultra (702B, 36B active) and GigaChat-3.1-Lightning (10B, 1.8B active) with a full DPO stage in FP8; the weights are available on HuggingFace and GitVerse under the MIT license. At the same time, the GigaChat assistant was moved to the flagship model with long-term memory across sessions.

Synapse runs AI agents across product, supplier, location, compliance, and performance data as a single digital twin, built on the multi-domain MDM core inherited from Riversand's 2021 acquisition. Human oversight of the agents is promised, not full autonomy.

Under the Yandex Commerce Protocol (YCP), stores on 1C-Bitrix and Yandex KIT, as well as Yandex Market sellers, can accept orders directly from search results or a dialogue with Alice without sending users to the website; the product must be included in the Yandex Products feed.

With Node.js 27 (October 2026), the release model changes: Alpha 6 months -> Current 6 months -> LTS 30 months. All releases become LTS, and the odd/even split is removed. The reason is minimal adoption of odd-numbered releases over 10 years and the volunteer team being overloaded by supporting four branches at once.

Existing APIs are wrapped into an MCP server without rewriting code, configured through API Manager and run on Flex Gateway. Part of Anypoint Platform, not a separate service.

Yandex AI Studio has made reasoning agents based on DeepSeek-V3.2 available to businesses (not a Yandex-developed model). DeepSeek’s weights are open from the developer, but Yandex access is available only through a closed cloud API. Agents solve multi-step tasks sequentially (tenders, code, automated tests), with cached/tool tokens, scenarios costing 4+ times less, a dedicated channel, and a no-request-logging mode.

Multi-year agreements with Accenture, BCG, Capgemini, and McKinsey to deploy Frontier platform agents in production at large companies. BCG and McKinsey provide strategy; Accenture and Capgemini handle integration with client data and cloud environments.

3.9.x is the last branch supporting ZooKeeper mode; there will be no minor 3.10 release. In this patch, KIP-1252 resolved the behavioral differences between ZK and KRaft. Since 4.0, Kafka runs only on KRaft; ZooKeeper was removed after 14 years of support, so companies using ZK mode should plan their migration.

The first release in the lineup with a .1 increment (previously .5): 77.1% on ARC-AGI-2, more than twice Gemini 3 Pro's score. Available through the API, Vertex AI, the Gemini app, and NotebookLM. Cloud only, with closed weights.

38 KIPs. KIP-932 (share groups/queues) has reached production-ready status. KIP-1147 unified CLI utility arguments (--bootstrap-server, --command-config) instead of the inconsistent legacy flags. KIP-1188 added an Allowlist policy for client configuration overrides in connectors, closing a vulnerability.

The release raised the minimum stack versions to PHP 8.3, MySQL 8.0.34 LTS, and Elasticsearch 8.17.0. It is part of the transition to calendar versioning (vYYYY.Minor) from the previous scheme.

The .NET 11 cycle begins: CoreCLR has been ported to WebAssembly, RISC-V and s390x support added, and the interpreter expanded. C# introduces Collection Expression Arguments. GA is planned for November 2026 (STS, not LTS).

The flagship Opus 4.6 added agent teams to Claude Code — agent teams divide large tasks into subtasks and work in parallel. Weights are closed; access is available only through the API/cloud (AWS, GCP, Azure).

A maintenance release with no new features - stabilization of the portal, business processes, connectors, and the configuration import/export mechanism. Upgrading from 3.1.2 or earlier requires intermediate migration steps.

Unauthorised remote code execution through the JobServer monitoring JMX port, CVSS 9.8. Patches: JobServer 8.0 TPS-6017 and 7.3 TPS-6018 (16 January 2026), Runtime 8.0.1.R2026-01-RT and 7.3.1-R2026-01 (24 January 2026). As a temporary measure, disable the monitoring JMX port or protect it with TLS authentication.

MCP server tools can return interactive HTML components, such as dashboards, forms, and multi-step wizards, rendered in an iframe sandbox directly in the chat with the model. The first extension to move from an SEP proposal (November 21, 2025) to official production-ready status.

Anthropic published a detailed document on Claude’s values and behavior — a public specification of principles instead of hidden system instructions.

Akeneo released its own MCP server to connect LLM agents to product data with access controls and guardrails, plus an integration with Stripe Agentic Commerce Suite for agentic checkout.

The new interface, featuring light and dark themes and vector icons, is available in beta alongside the legacy Taxi interface. The 1Cv8.1CD file database limit has increased from 6 to 16 GB.

The Odoo Enterprise Subscription Agreement (v12 terms) establishes the right to increase the renewal price by up to 7% for each year of the previous term if the current price is below the latest list price. Indexation now applies at every renewal rather than occasionally. In addition, from April 2026, a 25% surcharge applies to databases running versions older than the three latest major releases, for contracts renewed after 2025-07-04.

Reasoning mode based on technology that won gold at the IMO and ICPC: outperforms Gemini 3 Pro on HLE and GPQA Diamond. Available only with a Google AI Ultra subscription and through early access for select enterprises and researchers. Closed weights.

Yandex Cloud released AI Search: agents can access web search (with filters by language, region, and site) and corporate documents (PDFs and presentations) with source citations. Yandex positions this as a capability available worldwide from only a few vendors, such as OpenAI and Perplexity.

Protocol anniversary. Experimental task-based workflows were added (SEP-1686, statuses working/input_required/completed/failed/cancelled), Dynamic Client Registration was replaced by URL-based Client ID Metadata Documents (SEP-991), URL mode elicitation for OAuth was added without passing credentials through the client (SEP-1036), tool calling inside sampling requests was added (SEP-1577), and the tool name format was standardized (SEP-986).

Yandex B2B Tech opened access to Alice AI (LLM + ART), the new name for its flagship neural-network lineup replacing YandexGPT. According to the vendor, Alice AI LLM outperforms DeepSeek V3.1 and Qwen3-235B in 60% of business tasks. Available only through Yandex Cloud AI Studio, with closed weights.

The flagship for coding, agents, and computer use at release; new tools for long agent sessions and integrations with Excel, Chrome, and desktop environments. Pricing reduced to $5/$25 per million input/output tokens. Closed weights, available only through API/AWS/GCP/Azure.

Standardization of UI resource declarations, their binding to tools, and bidirectional messaging between embedded UI and the host. Initially limited to text/html in a sandboxed iframe. One of the community's most requested features, building on earlier MCP-UI work and the OpenAI Apps SDK.

New major release: the pipe operator (|>) for chaining calls without intermediate variables, URI extension (RFC 3986 + WHATWG URL, replacing parse_url), clone with for immutable objects, array_first()/array_last(), the #[NoDiscard] attribute, static closures and first-class callables in constant expressions, and backtraces for fatal errors.

In its Autumn Release 2025, Akeneo introduced a built-in DAM with AI asset auto-tagging, AI transformations (cropping and background editing), and a global CDN. Previously, DAM capabilities required external systems or middleware.

A family of 5 models featuring adaptive reasoning (deciding whether to think before answering) and configurable personalities. Codex-Max is designed for agentic coding. Available to everyone, with Pro/Plus/Go/Business receiving expanded access first. API/ChatGPT only, with closed weights.

Added Enhanced Content Layout Design for auto-generating rich-media item pages, AI GoPilot for extracting attributes from photos and validating data quality, and AVMT for automatically mapping brand attributes to a specific retailer's fields. The 'Is it Live' analytics were expanded to track publishing across more retailers.

A release with three years of support through 2028-11-10. C# 14: field-backed properties and extension properties. AI features: Microsoft Agent Framework and unified Microsoft.Extensions.AI abstractions are built into the platform. EF Core adds vector search for SQL Server 2025 and full-text search for Cosmos DB.

MuleSoft embeds ready-made integration components directly into the Salesforce Flow builder, without switching to Anypoint Studio or writing Mule configuration files. The target use case is no-code/low-code automation for admin roles, not integration developers.

Node.js 24.11.0 entered Active LTS on October 28, 2025, with support through April 2028. AsyncLocalStorage now uses AsyncContextFrame by default, URLPattern is globally available, and npm 11 is included.

Agent Builder is a native chat interface for Elasticsearch data and a builder for AI agents. DiskBBQ is a vector index that searches clusters directly on disk without loading them into memory (under 20 ms with 100 MB of memory). ES|QL now includes Smart Lookup Joins, matching across multiple fields through remote clusters.

Current release: the --allow-net flag gives explicit control over network access in the permission model; Web Storage (localStorage/sessionStorage) is enabled without a flag; V8 is updated to 14.1 with faster JSON.stringify and built-in Uint8Array base64/hex conversion.

Elastic closed its acquisition of Jina AI, adding embedding models, rerankers, and small language models for multimodal and multilingual search. Elastic confirmed that the models will remain open-weight, be published free of charge on Hugging Face, and also be available through Elastic Inference Service.

Free-threading (a build without the GIL) is officially supported, per PEP 779, and the penalty for single-threaded code has been reduced to about 5-10%. The JIT compiler remains experimental (PYTHON_JIT=1). The new t-strings syntax (PEP 750) provides template strings with deferred interpolation for safe SQL/HTML construction.

Visual Agent Builder (a drag-and-drop canvas for multi-agent workflows), Connector Registry for managing access to data (SharePoint, Teams, Drive), ChatKit for embedding chat agents, and enhanced evals. Cloud/API only, with closed weights.

An agent for building, managing, and scaling Mule applications through natural-language commands is built into Anypoint Code Builder and available free of charge to all Anypoint Platform users. It is a closed proprietary assistant that operates only within the MuleSoft ecosystem.

A family of safety classifiers (0.6B/4B/8B) for real-time moderation of LLM traffic: Gen (instruction-following) and Stream (token-level). Apache 2.0, open weights.

Odoo 19 was presented at Odoo Experience 2025 in Brussels: AI Server Actions enable no-code automation of fields and content from a text prompt, with ChatGPT/Gemini account connections for generation. Accounting features OCR for bank statements with manual correction, WhatsApp delivery notifications and ISO20022; HR introduces the new Equity module and Pay Runs instead of batch payroll processing.

KIP-932 (Queues for Kafka, share groups) entered preview and is not production-ready. KIP-1071 added the Streams Rebalance Protocol based on the new consumer group protocol (KIP-848), available through early access. KIP-877 gave plugins and connectors a unified metrics registration mechanism.

EmbeddingGemma-300M (open weights, Apache 2.0) ranks at the top of multilingual MTEB among models under 500M, with a 2,048-token context and Matryoshka embeddings truncated from 768 dimensions. In the medical domain, fine-tuning outperformed the larger Qwen3-Embedding-0.6B: 0.8862 versus 0.8493 NDCG@10.

Syndigo, Riversand's parent company, acquired 1WorldSync, including PowerReviews, adding GDSN data-pool synchronization and user-generated content such as reviews to its MDM/PIM stack. The deal was backed by Summit Partners, TJC, and Battery Ventures; the combined base has 18,000+ customers in 60 countries.

An image editing model based on Qwen-Image (native text rendering); open weights, self-hosting possible.