StepFun 3.7 Flash: A 196B-Parameter MoE Model That Runs Locally and Rivals GPT 5.5

China's StepFun (階躍星辰) has released Step 3.7 Flash. It uses a Mixture-of-Experts (MoE) architecture with 196B total parameters, of which only 11B are active during inference. The model can reportedly run locally on machines with 128GB of RAM.

The benchmark results are the most notable part. It scored 56.26% on SWE-Bench Pro, outperforming DeepSeek V4 Flash (55.6%) and Gemini 3.5 Flash (55.1%). On DeepSearchQA, it posted an F1 score of 92.82%, coming within striking distance of GPT 5.5's 93.98%. It also achieved 47.2% on HLE (tool use).

Given how lightweight it is with only 11B active parameters, its cost-performance for agent workflows and coding tasks is exceptional. It is also available on OpenRouter and NVIDIA NIM, providing options beyond self-hosting.

Glean Surpasses $300M in Annual Revenue: Enterprise AI Search Bets on Cost Reduction

Glean, an enterprise AI search platform, has surpassed $300 million in annual recurring revenue (ARR), achieving 3x year-over-year growth.

That growth is noteworthy at a time when tech giants are moving into the same space. Glean's strength lies in making "AI-driven cost reduction" its primary sales pitch, which appears to align with enterprises' growing need to optimize AI budgets. The repositioning—from traditional search efficiency to cutting AI spending across the entire organization—seems to have paid off.

Microsoft's Internal Data Suggests: Can AI Cost More Than Hiring People?

A report analyzing Microsoft's internal data has sparked debate with a surprising conclusion: using AI can potentially cost more than hiring human talent.

At a time when the cost-saving benefits of AI adoption tend to be one-sidedly emphasized, this analysis offers an important perspective. It argues that once you factor in actual operating costs, licensing fees, infrastructure investment, and the human review processes required to guarantee output quality, the simple equation of "AI = cheap" does not always hold.

That said, the conclusions can change dramatically depending on the analysis's assumptions and what is being compared, so caution is warranted against definitive interpretations.

Law Firm Kirkland & Ellis to Invest $500M in Proprietary AI System

Kirkland & Ellis, a major US law firm, has revealed that it will invest $500 million in building a proprietary AI system to support client work end-to-end.

This is believed to be one of the largest single-company AI investments in the legal industry. Rather than adopting general-purpose LLMs, the firm plans to build a proprietary system specialized for legal work, with the aim of using AI to cover a wide range of tasks from contract review to due diligence.

The legal industry deals with enormous volumes of document processing, making it one of the fields with the most to gain from AI-driven efficiency. Kirkland's investment can be seen as a symbol of AI adoption getting serious in specialized industries.

Ohio Halts Data Center Tax Incentives: Headwinds for AI Infrastructure

Ohio has ended its tax incentive program for data centers. As tech companies accelerate the construction of AI data centers, the policy shift comes amid concerns about the growing burden on local communities.

According to reporting by the AP, the state had grown concerned about the impact of the rapid proliferation of data centers on power demand and local infrastructure. The rollback of tax incentives suggests that scrutiny of the social costs of AI infrastructure investment is intensifying.

Similar debates could emerge in other states and countries, and how the benefits of AI industry growth are shared with local communities is expected to become a key policy issue going forward.

NBA to Introduce AI-Powered Automated Calls

NBA Commissioner Adam Silver has announced the introduction of an AI-based automated system for out-of-bounds calls.

While AI in sports has centered on video analysis and tactical analysis, directly applying AI to actual in-game officiating is a major step forward. As a technology that supplements—or even replaces—human referees' judgment, spillover effects to other sports are anticipated.

Improvements in the objectivity and accuracy of calls are expected, but operational challenges remain, including the transparency of AI decisions and mechanisms for disputing them.

Fujitsu Announces Partnerships with OpenAI and Anthropic on the Same Day

On May 28, Fujitsu announced partnerships with OpenAI and Anthropic in quick succession. Attention is focused on why Fujitsu, which develops its own AI technology, is teaming up with external AI vendors.

The move appears to be a strategy that leverages the two companies' distinct strengths: combining OpenAI's broad ecosystem and Anthropic's focus on safety with Fujitsu's enterprise customer base. A major Japanese IT company partnering with multiple AI providers at once signals that AI adoption is becoming increasingly diversified.

---