September 17, 2026 — The Trend Tribune

By tech_tanaka · September 17, 2026

Without a blog post or press release, the 744B model was unveiled—the contents of "Atria Dawn" quietly released by the Shanghai AI Research Institute. — (Photo: illustrative)

The Release of the 744B Model Without a Blog Post or Press Release

On September 11th, a repository quietly appeared on Hugging Face: "Atria-Dawn-Preview." It was posted by the Shanghai AI Laboratory, a Chinese state-run research institution. It's a 744 billion parameter MoE (Mixed Expertise) model, with all its MIT-licensed weights publicly available. There's no blog post, no press release. No price list, no API information. All that's there is a model card and approximately 1.5 terabytes of weight files. This quiet release, occurring at a time when major Western labs are publicly debating the slowdown in AI development, piqued my interest as an engineer.

Built on GLM-5.2, Crafted by Another Lab

What's technically interesting is the origin of this model. Atria Dawn Preview is based on the GLM-5.2 MoE (Mobility of Environment) platform model with 744 billion parameters, released under the MIT license on June 13th by Z.ai, a subsidiary of Zhipu AI. With approximately 40 billion parameters activated per token, it's designed to be offered at an exceptionally low price for a frontier-level model. The Shanghai AI Research Institute used this same GLM-5.2 platform, applying their own post-training methods to create a separate product called "Atria Dawn." In other words, two different research institutions refined the same silicon (the same platform model) in different directions, releasing one as a general-purpose chat coding model and the other as an agent model specialized for research and experimental tasks.

What was this model created for?

The Atria Dawn Preview model card describes itself as an agent model designed to "lead research questions to executable, verifiable, and reproducible results." It is designed for research and engineering applications that require continuous environmental understanding, tool utilization, and the completion of multi-step tasks, handling a series of loops including problem analysis, solution design, tool utilization, code implementation, experiment execution, results analysis, and recovery from failure. The context window has 256,000 tokens and is text-only; it does not support image input. It appears that the decision was made to discard the 1 million tokens of context and image support that GLM-5.2 has, and allocate those resources to the ability to execute the research loop.

Benchmark Details and Unverified Aspects

The paper accompanying the model card claims that it demonstrated performance that competes with frontier-class agents in 16 benchmarks spanning "real-world research, engineering, and digital work," and achieved the highest score in five of them. However, the abstract does not reveal the specific names of those five benchmarks or the details of the score differences. Another interesting point is the analysis of human-agent collaboration. The paper describes how "the agent frequently proposes methods and implements modifications, while humans retain most of the final judgment, guiding the exploration through decision-making and feedback," positioning this as a "shift from task-level execution to project-level partnership." However, it's important to note that these claims are based on self-reported benchmark results and have not been verified by a neutral third-party organization.

Why the "Quiet" Release?

The very avoidance of a flashy announcement is seen as a signal. The Shanghai AI Research Institute is a lab with a track record of consistently releasing models like Intern-S2 and InternLumina this year. It's still unclear whether Atria Dawn is the first step in a new development in that series or a one-off experimental release. However, the fact that the comparison table on the model cards uses GLM-5.3 (its successor) instead of GLM-5.2 as the comparison target is noteworthy as an editorial effort to improve the visual appeal of the open weight column.

Things Developers Should Note

From an engineer's perspective, the approach of retraining the same foundational model for different purposes is itself insightful. Whether refining it as a general-purpose chat model or as an agent capable of handling long research loops—this release clearly demonstrates that even with the same foundation, the design of post-training can result in completely different products. While the benchmark claims themselves cannot be taken at face value as they have not yet undergone third-party verification, the strength of open weight models lies in the fact that the weights are fully publicly available under the MIT license, allowing for hands-on testing and verification. We will be closely watching how independent evaluation sites position this model in the coming weeks.

References: huggingface.co / www.orcarouter.ai / startupfortune.com

オープンソースLLM上海AI研究院AIエージェント中国AI

"We won't be going public this year"—An accountant analyzes two contrasting IPO strategies: OpenAI's postponement and Anthropic's progress.

By fin_suzuki

CEO Sam Altman has denied that OpenAI will list in 2026. This article, from an accountant's perspective, analyzes the underlying concerns about the stock price volatility following SpaceX's IPO, as well as the contrast with Anthropic, which moved forward with a Nasdaq listing in the same week, all while highlighting the company's financial strength ($122 billion raised and a $4.7 billion credit line) and its emphasis on safety.

A Clear Statement: "We Will Not Go Public This Year"

OpenAI CEO Sam Altman stated in an interview with Fortune magazine that the company will not go public in 2026. He said, "Going public at this moment would be an ill-advised moment." This comes as a reversal of policy, as the company had already submitted a private S-1 filing to the SEC in June, and there were initial speculations that an IPO application might be filed as early as September. Considering that Anthropic informed investors it was profitable in preparation for a Nasdaq listing (as previously mentioned), these two frontier AI companies are showing contrasting moves in their IPO strategies. As an accountant, I would like to break down the details of this "postponement" using numbers.

Listing the Numbers

In March of this year, OpenAI raised $122 billion in the largest single private funding round in history, bringing its post-investment valuation to $852 billion. The joint lead managers are SoftBank ($30 billion) and Amazon ($50 billion), and Amazon has also been designated as OpenAI's exclusive third-party cloud partner. It is said to have 900 million weekly active users and annualized revenue exceeding $20 billion. While some speculated that the IPO would result in a valuation of $1 trillion, this statement pushes that date back to at least 2026.

Official Reasons and Other Factors

Altman cited "security" as his reason. He stated that "it's premature to go public now, given the ongoing discussion in the AI industry about slowing down the pace of capability improvement." In the same week, Anthropic CEO Dario Amodei also published an article arguing that "the pace of AI model capability improvement should be slowed." While both CEOs are unequivocally calling for an industry-wide slowdown, Anthropic is emphasizing profitability and moving forward with its Nasdaq listing, while OpenAI is postponing its IPO—this contradiction is difficult to explain solely by the ostensible reason of security concerns. According to reports, the New York Times reported in June that "seeing the volatile stock price fluctuations after SpaceX's IPO (rising to a valuation of $1.8 trillion after the initial price before plummeting) had led OpenAI to reconsider its IPO timing," suggesting that caution regarding the market environment itself cannot be ignored as a contributing factor to the decision.

Reasons for Continuing Fundraising While Remaining Private

What's noteworthy from an accountant's perspective is OpenAI's explicit statement that it "does not feel a financial need to rush into an IPO." In addition to a massive $122 billion in funding, it has reportedly secured a $4.7 billion revolving credit line, indicating that it has the financial strength to manage its immediate cash flow even without the liquidity that would only be gained through an IPO. Remaining a private company also has the advantage of allowing decision-making without being constrained by the short-term interests of shareholders. Altman himself has stated that "it is necessary to be able to make decisions that do not necessarily align with the short-term interests of the business or shareholders," a reasoning that is typical of a company that has maintained a hybrid non-profit/for-profit structure.

Contrast with Anthropic

As discussed previously, Anthropic has informed investors that it expects to post a profit in adjusted operating income for the second quarter, and has progressed to the point where there are even speculations that its valuation will exceed $2 trillion in preparation for a Nasdaq listing. While both are frontier AI companies, one is accelerating its IPO preparations with profitability as its goal, while the other is postponing its listing backed by its massive financial resources—this difference seems to reflect not only the differences in their business models (gross profit margins of APIs and customer base), but also the differences in their management philosophy regarding the choice of going public itself. The decision of whether to prioritize the fundraising power gained through an IPO or the freedom of decision-making that comes with remaining private is an interesting crossroads in predicting the future of corporate governance for AI companies.

What Accountants Should Watch

The statement that they will not list in 2026 does not guarantee a listing in 2027. Altman did not specify a concrete next date, only stating that "there is still much to be done regarding safety and alignment." The private S-1 draft submitted in June can be interpreted as a strategic move to avoid a concrete commitment while keeping the option of listing open. What should be watched closely going forward is how long this private strategy, backed by massive financial resources, will continue, and how the actual listing of Anthropic will influence OpenAI's decision.

References: fortune.com / techcrunch.com / www.ibtimes.com

OpenAIIPOファイナンスSam AltmanAnthropic

What do videos showing "robots attacking humans" fail to tell us? — An examination of the Unitree incident in Zhejiang Province.

By robo_takahashi

A video of a Unitree humanoid robot kicking an engineer, filmed in Shaoxing, Zhejiang Province, has gone viral. While much of the media coverage has described it as a "warning about AI safety," the fact that the robot was actually being remotely controlled is often overlooked. Taking into account the "fighting demonstration" culture of the Chinese robotics industry, this article will examine the theory of runaway autonomous AI versus a malfunction of remote control.

15 Seconds Filmed in Zhejiang Province

On September 7th, surveillance camera footage from a facility in Shaoxing, Zhejiang Province, went viral on social media. The 15-second video shows an engineer operating a humanoid robot with a remote control when the robot suddenly turns towards the engineer, raises its leg, and knocks the remote control away. Headlines such as "Robot Attacks Human" and "Warning about AI Safety" were picked up by media outlets worldwide, and the video continued to spread as of September 15th. As someone involved in robotics, the first thing I want to confirm whenever this kind of footage appears is, "Is this truly a case of an autonomous AI malfunctioning, or is it something else?"

The First Fact to Understand—It Was a Unitree Robot and Was Being Remotely Controlled

Crossing multiple news reports reveals the following facts: The robot in question was a Unitree humanoid robot, and the incident occurred while an engineer was using a remote control to activate a program and conduct a test. In other words, this isn't a case of "an AI making autonomous decisions going rogue," but rather a case of "a remotely controlled robot performing an unintended action." This distinction is extremely important. The former raises security concerns for large-scale language models and behavioral planning AI, while the latter is closer to traditional robotics challenges such as control system bugs, communication malfunctions, or program misfires. Much of the reporting is framed as "concerns about AI security," but based on the information currently available, there is no way to definitively determine whether it was caused by the AI's autonomous judgment or simply a control system malfunction.

The Context of "Robot Fighting" Culture Cannot Be Ignored

Another point worth noting is that the Chinese humanoid industry already has a well-established culture of showcasing technological prowess by having robots perform martial arts-like movements. In December 2025, EngineAI released a video of its T800 robot kicking its CEO, explaining that the purpose was to prove it wasn't CGI. In May of the same year, China Central Television (CMG) hosted a "Fighting Robot Competition" in Hangzhou, and footage of Unitree's G1 robot performing boxing-style moves garnered attention. In June, an incident was reported where a robot kicked a child during a performance at an amusement park. Therefore, before concluding from this video alone that "the humanoid robot autonomously displayed hostility towards humans," it's necessary to distinguish between this "intentionally aggressive behavior" and "unintentional malfunctions."

Why the Speed and Closeness of the Video Arouse Concerns

One reason this video attracted particular attention is the speed of the robot's movements and its physical proximity to the engineers. The fact that robots can now perform complex human-like movements at high speed and with great force, right next to humans, is undeniable evidence of the progress made in humanoid robot development in recent years. However, progress and safety need to be evaluated on different axes. The extent to which safety protocols were in place for testing in environments close to humans—such as emergency stop mechanisms, physical safety fences, and limitations on the range of motion—cannot be gleaned from this report.

What Engineers Should Consider

Every time this kind of video goes viral, there's a temptation to jump to the "AI attacked a human" narrative. However, from the perspective of someone familiar with robot development, I believe we should start by considering the fact that it was being remotely controlled. Rather, what this incident truly raises is not the danger of AI autonomy, but a more mundane but fundamental issue: the safety design when humans test robots in close proximity—how much of a mechanism is incorporated to stop unintended actions before they reach humans, even when remotely controlled. The fact that the IEEE International Conference on Humanoid Robotics is focusing on "Occupational Safety and Certification" this year is evidence that the industry is beginning to seriously acknowledge this issue. Without being swayed by sensational headlines, we should keep a close eye on whether Unitree and related organizations will provide a technical explanation of the cause.

References: organiser.org / www.ntd.com / ca.news.yahoo.com

Unitreeヒューマノイドロボット安全性中国ロボティクス遠隔操作

Without a blog post or press release, the 744B model was unveiled—the contents of "Atria Dawn" quietly released by the Shanghai AI Research Institute.