On July 15, U.S.-based Google announced a major feature improvement update for its open model “Gemma 4.” Reflecting community feedback, the update centers on faster inference processing and various bug fixes, creating a more developer-friendly environment.

The core of this update is the activation of “Flash Attention 4 (FA4)” support on NVIDIA’s latest GPU architecture, “Hopper.” This reduces the time to first token (TTFT) by up to 31%. Furthermore, throughput during the prefill stage, which preprocesses large amounts of data, has improved by 25% to 70%, significantly enhancing overall response performance.

On the functionality side, the chat template has been improved, enabling more natural and smoother conversational responses. The behavior when calling external tools and APIs has also been corrected, promising more accurate and consistent execution.

Image recognition capabilities have also been strengthened. The default vision token count is set to 280, prioritizing token efficiency, but by raising the setting to a maximum of 1120, optical character recognition (OCR) accuracy improves, and resolution can be increased to the equivalent of 2.51 megapixels to read fine details. A demo of this new feature is publicly available on the AI development platform “Hugging Face” Spaces.

Gemma 4 is an open model Google released on March 31 of this year in four sizes: E2B, E4B, 31B, and 26B A4B, adding the 12B Unified version in June. This updated version is available from the Gemma 4 collection on Hugging Face.

Microsoft Discloses Windows 11 Malfunction on Dell PCs

On July 14, Microsoft disclosed a known issue with Windows 11 that could cause performance degradation on certain Dell devices.

The affected devices are some Dell machines that installed the preview update (KB5095093) released on June 23. In this environment, the “Intel Innovation Platform Framework Processor Participant” driver in Device Manager displays a yellow warning (exclamation mark), potentially causing adverse effects on overall system performance, power consumption, and operation.

Microsoft has identified the cause as an incompatibility between the Intel driver and a new “Windows USB-C Connection Manager” interface introduced in the update and is working with Dell on a fix. As a temporary mitigation, the monthly security update (KB5101650) that began rolling out on July 14 will not be offered to affected devices until the issue is resolved. The company plans to provide a formal fix within days. The impact scope covers Windows 11 versions 25H2 and 24H2; Windows Server is not affected.

Meta Unveils “Muse Spark 1.1,” Enhancing Autonomous AI Agents

On July 9, Meta announced “Muse Spark 1.1,” the latest multimodal reasoning model developed by Meta Superintelligence Labs. This model is specifically designed for AI agents executing complex tasks, with significantly enhanced capabilities in tool utilization, direct computer operation, coding, and multimodal understanding compared to the first-generation model.

It can autonomously manage a massive context window of up to 1 million tokens, formulate plans as the main AI agent, and then delegate parallel execution to multiple sub-agents. This enables it to handle long, complex workflows spanning multiple applications and services. For computer operation, it has been trained to choose between creating automation scripts and direct manipulation, such as clicking like a human, depending on the situation. In coding, it handles bug fixes, new feature implementation, and code migration in large codebases, demonstrating performance comparable to major competing models in Meta’s internal benchmark tests.

Simultaneously, a public preview of the “Meta Model API” for developers was launched. General users can use this model in “Thinking” mode on the Meta AI app and meta.ai.

OpenAI Announces “Codex Micro” Macro Pad for Codex

On July 15, OpenAI announced the “Codex Micro (kbd-1.0-codex-micro),” a compact input device dedicated to its coding agent “Codex,” co-developed with keyboard manufacturer Work Louder. It is sold on the company’s official merchandise store “OpenAI Supply Co.” for $230.

The device features 13 mechanical switches, plus a touch sensor, rotary encoder, and joystick. Each key supports status display via RGB LEDs linked to Codex, allowing users to see at a glance whether multiple agents are in a “Processing,” “Running,” “Waiting,” or “Completed” state without switching chat screens. Flicking the joystick can launch workflows such as pull request (PR) reviews, debugging, and refactoring, while approval, rejection, push-to-talk, and starting a new chat can be assigned to dedicated keys. Turning the dial allows on-the-fly adjustment of the AI’s reasoning level. Connectivity supports Bluetooth and USB Type-C, and it works with both Mac and Windows.

Intel to Invest an Additional ¥812.5 Billion in Irish Facility

On July 13, Intel announced it would invest €5 billion (approximately ¥812.5 billion, or $5.0 billion) in its Leixlip campus in Ireland to expand semiconductor production capacity. This is in response to surging demand for AI and high-performance computing (HPC), and will increase production of the server CPU “Xeon 6” and next-generation Xeon processors manufactured on the company’s “Intel 3” process.

The investment centers on retrofitting existing factories and installing cutting-edge manufacturing equipment, including the expansion of an automated material handling system to integrate the various buildings scattered across the campus into a single production environment. This capital investment program has already begun in early 2026 and is expected to create specialized jobs in construction and equipment installation, as well as new full-time high-tech positions at Intel.

Since entering Ireland in 1989, Intel has invested a cumulative total of over €30 billion in the country. Leixlip is one of the company’s most advanced manufacturing sites, currently employing approximately 4,900 people. The announcement also cited contributions to the EU’s technological sovereignty and strengthening the European semiconductor supply chain as the significance of the investment, with Irish Prime Minister Martin and the CEO of government agency IDA Ireland expressing their welcome.

AI Industry’s Mixed Fortunes: TSMC Posts Strong Earnings, IBM Plunges

This week in the AI industry, corporate fortunes diverged sharply. Taiwan Semiconductor Manufacturing Co. (TSMC), the world’s largest contract chipmaker, reported explosive growth in its second quarter of 2026 earnings, with revenue up 36% year-over-year and net profit surging 77.4%, buoyed by demand for AI semiconductors. Full-year revenue has skyrocketed from approximately $75.99 billion in 2022 to approximately $122.56 billion in 2025. CFO Wendell Huang indicated that this strong demand will continue into the current quarter, as companies are seeking the maximum amount of AI chips they can procure.

Conversely, on July 15, IBM pre-announced its second-quarter results, revealing that revenue and earnings per share (EPS) fell significantly short of analyst expectations. The primary cause is a shift in customer spending from the company’s mainframe systems to AI servers and memory/storage chips. Against analyst expectations for adjusted EPS of $3.02 and revenue of approximately $17.86 billion, actual adjusted EPS was $2.93 on revenue of approximately $17.2 billion. Following this announcement, IBM’s stock price plummeted more than 25% on the same day, marking its steepest drop since 1968, and ended the week down 26%.

Ashish Nadkarni of market research firm IDC suggested the market reaction was somewhat excessive but noted, “It means that even IBM is not immune to the strategic reallocation of corporate budgets to accelerate AI adoption.”

The Philadelphia Semiconductor Index (SOX) also fell about 10% for the week, entering bear market territory. NVIDIA dropped roughly 4% for the week, while Intel was sold off heavily, falling 13%. Despite TSMC’s strong earnings, concerns are spreading in the market about the sustainability of AI investment. The next major catalysts will be the earnings reports from Google and Intel, scheduled for the fourth week of July.