Freshworks acquires Israeli AI startup Robusta Dev

Itaú Unibanco announced on Monday, the 26th, the acquisition of NeoSpace AI, a company from Uberlândia, located in the Triângulo Mineiro region. The deal involved a value of R$ 2.925 billion for the startup, which specializes in creating foundational artificial intelligence models.

NeoSpace AI had already maintained a collaboration with the financial institution for almost two years. During this partnership period, the technology developed by the startup was applied in various areas of the bank, such as credit analysis, fraud prevention mechanisms, and improving customer relationships.

For the transaction to be finalized, Itaú informed that it still requires approval from both the Administrative Council for Economic Defense (Cade) and the Central Bank of Brazil (BC).

With this purchase, the bank now possesses an internal structure dedicated to developing sophisticated artificial intelligence models specifically focused on the financial sector. Itaú highlighted that this will enable the creation of more customized solutions, which consider the moment and context of each client when assisting in their financial decisions, always maintaining rigorous care for security, privacy, and responsible data use.

Furthermore, the operation strengthens the innovation ecosystem of the city of Uberlândia. According to data collected by UberHub, the startup community of Uberlândia, the city's technology companies have generated about R$ 13 billion in investments, acquisitions, and added value over the last eight years. With the inclusion of NeoSpace AI, this total value approaches R$ 16 billion.

According to Itaú, NeoSpace is dedicated to developing foundational artificial intelligence, which focuses on identifying complex patterns and relationships in large volumes of data. This specialization allows the startup to create models tailored to the specific characteristics of different market segments.

The bank detailed that the development of these models expands the options for personalizing banking services, offering solutions capable of addressing different financial situations and needs throughout the customer relationship, respecting the principles of security, privacy, and ethical data use.

Additionally, Itaú emphasized that the integration of the startup represents a natural progression of a partnership that already showed good results and which now gains the conditions to boost new projects in the field of artificial intelligence.

It is important to note that, even after the acquisition, NeoSpace AI will maintain its operations autonomously. The company will remain under the management of CEO Bruno Pierobon, preserving its team and the research culture established since its founding.

DeepSeek has released DeepSeek Harness version v0.2.1-rc1 on October 17th. This version is the first candidate release of the 0.2.1 series for its open agent runtime. The release notes on GitHub compile the main changes concerning users and developers since the release of v0.2.0-rc.2. Some of these improvements, including reasoning trace translation, the Git Worktrees plugin, and the Claude Code Mods compatibility layer, were previously presented in the alpha.1 and alpha.2 preliminary releases, as reported by Pandaily.

The main innovation is the expansion of plugin capabilities. Users can now find and install plugins from custom index sources, not just from the official list. The plugin manager has gained a 'Allow Agent to Create Plugin' function, which saves a draft in creation mode and only starts working after receiving a request from the user. Functions for enabling, disabling, deleting, and updating plugins now synchronize dependencies. With hot reload enabled, an updated or reinstalled replaceable plugin takes effect immediately; otherwise, the application requests a restart.

Additional plugins support minimal mode, while the official list allows installing Claude Code and Codex packages corresponding to the current release for headless profiles, SDKs, and ACPs. An experimental session state interface allows plugins to independently save and read their session records.

Progress was also made in model and context processing. Users can now set reasoning levels for each model via the interface instead of using a configuration file, utilizing custom model APIs. The pi-ai router manages system prompt updates within a session and dynamically adds or removes tools according to the model's capabilities. Minimal mode can read images and automatically compress long dialogues, while web and desktop applications have added the /compact command for manual compression. Global instructions can read the AGENTS.md file from the general agents directory, which is specified via the DSH_AGENTS_HOME environment variable.

A new working_directory tool has been introduced for developers, along with TypeScript and Python SDK interfaces for reading and switching the session working directory. The web version allows logging into DeepSeek via local HTTP, and secrets and tokens found in sent messages are masked by default. New sessions can be pre-filled with an unsaved prompt, and the Markdown preview displays the YAML header as a list of fields.

It is worth noting that an automated account on X reported the brief appearance of an entry labeled 'deepseek-v4.1-flash-1017' in the official list of DeepSeek API models before it was removed. The suffix corresponds to a dated naming scheme of earlier snapshots, such as V4-Flash-0731 and V4-Pro-0813. The same account noted that the public model deepseek-flash still points to V4.1 Flash, not build 1017. DeepSeek has not commented on the existence of such a snapshot or what it might change.