Tool news

Sber released GigaChat 3.5 Ultra: a 432B MoE model under MIT, FP8 training, online RL, and MTP acceleration. For enterprise, this is an on-prem/self-host option, not a new API tier.

The release of boxed versions of standard products is being reduced as distribution shifts to electronic delivery.

A joint project was announced to implement the DDMRP methodology (MRP 3.0) on 1C:Enterprise for supply management in unstable conditions. It is accompanied by a series of webinars based on 1C:ERP.

A new release of the standard 1C:ERP WE 2.5.27.52 configuration has been issued; related 2.5.27 releases updated the ERP/KA and UT stacks.

Compatibility has been confirmed for the sales management suite with the built-in CRM v3.1 features and the CRM module for 1C:ERP.Holding Management.

Three days after release, a U.S. export directive required Anthropic to disable access to Fable 5 and Mythos 5 for all foreign nationals inside and outside the country. Closed weights in a vendor cloud can be revoked instantly by regulators, leaving enterprises with no lever to preserve access.

Fable 5 is Anthropic's most powerful public model, SOTA on almost all benchmarks; Mythos 5 offers the same power without safety classifiers, available only in limited form through Project Glasswing. The weights are not published: access is only through Claude API, Bedrock, Vertex, and Foundry, with closed weights.

1C launched a training program for the mandatory transition to EPD from 01.09.2026; support is being implemented in sales and ERP configurations. A regulatory requirement in CIS.

The RC removes the initialize handshake and sessions: state is passed via explicit handles, follow-ups go through InputRequiredResult without SSE. Load balancing over plain HTTP without sticky sessions.

The RC introduces Extensions (reverse-DNS IDs, capabilities, versioning): MCP Apps serves HTML in a sandbox iframe, Tasks offloads long-running operations. A narrow core, growth through modules.

The RC carries W3C Trace Context in _meta for end-to-end trace correlation and introduces a lifecycle policy: at least 12 months between deprecation and removal.

2.4.9 added Symfony 7.4 LTS and PHP 8.5, extended CAPTCHA to REST/GraphQL (closing the API bypass). 581 core improvements; the clearCart mutation is now open to Open Source.

An MCP server was announced at Adobe Summit 2026: AI agents gain access to the catalog, cart, prices, stock and checkout. It depends on the Integration Starter Kit and AppBuilder.

The AI suite analyzes the implementation, builds a migration path to ACCS, and rewrites legacy code into event-driven components. Status is coming soon; it is not publicly available yet.

Studio (React/TypeScript) becomes the only interface, all ExtJS is removed, bundle versions are unified. The 2026.x branch requires Symfony 7.4.

AI agents read and propose changes to objects and assets through MCP under the same permissions and workflows as people. All actions go through the Data Spine layer.

deliveryOptionsCalculate replaces implicit shipping webhooks. Search has been redesigned (prefixes, operators, ranking); gift cards are paid via the Transaction API.

The roadmap sets 4 priorities (transport, agents, governance, enterprise) and a shift from a release-centric plan to Working Groups. Enterprise via extensions, not the core.

Saleor released Paper — a production-ready storefront with a mobile multi-step checkout. It includes an AGENTS.md with rules for AI tools (Cursor, Claude, Copilot).

Adobe announced support for the open UCP (Google) and ACP (OpenAI/Stripe) protocols, in addition to Agent Payments Protocol, so products can be purchased through ChatGPT and Gemini.

Share groups declared production-ready: consumers read the same partitions with per-record acknowledgment — a classic queue on top of Kafka.

Kafka Streams group management has been moved to the broker side: task assignment, topology storage, and observability are now generally available.

Kafka Streams exception handlers gained built-in routing of failed records to a dedicated queue (DLQ) instead of manual wiring.

Akeneo released a native MCP server: the catalog as resources, agent edits as tool calls, and server-side quality rules. The release also includes Digital Showroom and Custom Components

In its Winter Release, Akeneo announced a partnership with the Stripe Agentic Commerce Suite: product discovery by AI agents and acceptance of agent payments. The release's MCP server is covered separately

Odoo detailed its Indonesia localization package: Coretax XML e-Faktur, native Shopee, Xendit, and InterActiveQRIS integrations, plus tax adjustments. For ERP projects this signals that localization is becoming an architectural decision, not an after-the-fact add-on.

The final 8.5.1 release of the 8.5 line is out, with a new interface (light/dark theme, beta status within the release) and backward compatibility with 8.3. UNF 3.0.13, Retail 3.0.13, and Cashier Workstation 1.1.1 were updated at the same time.

The partnership combines POS ordering, a payment terminal, and backend management in Odoo for retail and F&B in Hong Kong. In practical terms, Odoo continues to strengthen scenarios where ERP, POS, and payments must work as a single operational flow.

The last release with Admin UI Classic alongside Studio. An experimental MCP server reads data over Elasticsearch/OpenSearch. Symfony 7 alongside 6.4, with 3-year LTS.

Odoo reports that mandatory Veri*Factu deadlines shift to January 2027 for companies and July 2027 for self-employed. Veri*Factu modules have been available in Odoo since July 2025 from version 17 onward — relevant for planning the compliance layer of ERP projects.

Instead of the problematic DCR (SEP-991) — Client ID Metadata Documents (client_id as a URL pointing to JSON). Cross App Access (SEP-990) shifts authorization to the organization's IdP.

Saleor shipped an official Adyen integration for international payment acceptance and refunds. It is implemented through the App model rather than a built-in core plugin.

Odoo wrapped up its 2025 Partner Awards at Brussels Expo, recognizing integrators across Europe, the Americas, Africa, and MENA. For buyers this is not a product feature but an indicator of the platform's implementation network maturity.

An MCP server launched, giving LLMs access to store data (channels, orders, products, stock) over Streamable HTTP. Read-only, no mutations; Saleor 3.21+ and Saleor Cloud.

The app connects a Saleor store to ChatGPT via OpenAI ACP, exporting the catalog to an ACP feed. For now the feed is exported manually; payment and checkout in ChatGPT are the next stage.

GenAI in Akeneo fills numeric and Yes/No attributes, extracts specifications from PDFs with unit conversion, and pulls values from reference entities. AI Configuration has been added

The rules engine gained triggered rules — real-time execution on product updates, tied to workflows. Akeneo DAM was re-released with AI asset tagging

Copilot generates text through Chat Completions, so it connects to different LLM providers (OpenAI, Hugging Face, RunPod). Documents appears in Studio for the first time (preview).

At Odoo Experience, the company revealed its 2025 target: EUR 650 million in billing revenue, 42% ARR growth, and a plan for 10,000 employees in 2026. Product highlights include payroll, industry packages, and stronger HoReCa/POS scenarios.

Odoo 19 adds AI summaries to CRM, AI sorting rules in Documents, a redesigned quoting flow, POS presets, and expanded industry scenarios. The release positions Odoo as an operational platform rather than a collection of ERP modules.

A public catalog and API of MCP servers opened at registry.modelcontextprotocol.io (Anthropic, GitHub, PulseMCP, Microsoft). Preview, with no data guarantees and breaking changes until GA.

Release 4.1.0 brought share groups to preview (not for production) and added early access to the Kafka Streams rebalancing protocol (KIP-1071).

GPT-5 launched as a cloud model: the weights are not published, and access is through the API and OpenAI products. The top tier remains closed, increasing dependence on the vendor cloud.

DeepMind showed Genie 3, generating interactive worlds; access is limited to a preview, and the weights are closed. It demonstrates a move toward closed research models.

OpenAI released two open-weight models under Apache 2.0 — downloadable and self-deployable. These are a separate line: GPT flagships remain closed; only these models are open.

Akeneo introduced its first AI agent, Data Architect Agent: based on sample product files and context, it drafts a data model. GA for Product Cloud is July 12, 2025

PX Insights shows how AI assistants like ChatGPT interpret and recommend products. In Spring 2026 — a closed loop that suggests missing words and attributes

Over 30 Symfony components now run in version 7 alongside 6, paving the way for migration. Asset metadata select fields pull values from external sources; the Datahub File Exporter writes to the directory tree.

The multi-tenant SaaS edition of Adobe Commerce became generally available (Americas from June 25, 2025, then EMEA and JPAC). Versionless with automatic Adobe updates instead of self-hosted/PaaS.

Anthropic released the Claude 4 line (Opus and Sonnet) through API and cloud providers. The weights are not published, so the model cannot be deployed in your own environment; privacy and availability depend on the vendor.

3.21 added the first OpenTelemetry stream (traces and metrics) for APM tools (Datadog, Grafana, Uptrace). It covers Saleor Core, apps and services.

Alibaba released Qwen3 with open weights under Apache 2.0, including MoE variants. Unlike Western flagships, the strong model can be downloaded and deployed in your own environment.

2.4.8 fixed ~497 core bugs, added PHP 8.4, MySQL 8.4 and MariaDB 11.4 LTS, extended GraphQL for headless and B2B (negotiable quotes), and strengthened 2FA with Duo Security.

Meta released Llama 4 with open weights (community license) and MoE architecture — deployable on-premises. Open-weight remains Meta's bet against closed-model flagships.

Gemini 2.5 Pro has improved reasoning and long-context handling, but the weights are not published, and the model is available through Google AI and Vertex AI. The top-tier Gemini remains closed.

The first major release entirely without ZooKeeper: metadata is managed by default by the built-in Raft (KRaft), reducing operational complexity.

The GigaChat 2 lineup expanded context and quality. For CIS enterprise, data processing in CIS data centers and ruble pricing matter more than open weights: the flagship remains cloud-based.

Google released Gemma 3 with open weights for self-hosting, but it is not a flagship. The strongest Gemini remains cloud-based and closed; the open-weight role was given to the lighter family.

Yandex published a lighter version of YandexGPT 5 Lite with an open pretrain checkpoint, but it is a lightweight model, not a flagship. The senior models remain cloud-based and closed.

Yandex updated its flagship YandexGPT 5 Pro in Yandex Cloud: API access, processing in CIS, ruble billing, and claimed compliance with GDPR. The senior model weights are closed.

Sber introduced GigaChat MAX as the top-end version of the line. The model runs in a CIS cloud with ruble billing and claimed compliance with GDPR; the flagship weights are closed.

The updated Claude 3.5 Sonnet and computer use mode have expanded agent workflows, but the model is still available only as a service. For GDPR, a privacy gateway is needed, not local deployment.

Qwen2.5 established a consistent open-weight lineup across sizes from lightweight to large — a solid base for on-premises deployment with full data control and no cloud dependency.

The o1 family focuses on step-by-step reasoning, but it is delivered only as a service, with closed weights. Strong reasoning without the right to local deployment.

Llama 3.1 405B demonstrated that open-weight models can compete with closed flagships. Self-hostable, eliminating vendor cloud dependency when handling sensitive data.

Anthropic has never published the weights of its models, and the entire Claude line is delivered as a cloud service. That locks in vendor dependence when working with sensitive data.