I witnessed the "darkest hour" of MiniMax

At the most intense moment, internal quarrels broke out as expected. After things calmed down, I chatted with the two main parties involved in the dispute. "I knew he was right back then, but I just couldn't help but argue with him in the heat of the moment."

"At the end of the day, there was no personal grudge between us after the fight. We were just bickering over extremely granular internal parameters." I later learned that on a separate occasion, when the pricing controversy was revisited, he directly defended the colleague he had called out, even though that person was not present at the discussion.

"What he lacked was context. That's everyone's responsibility."

Preface:

In the past June, I gained access to the inner workings of the large language model company MiniMax, with unusually high permissions that allowed me to witness nearly the entire launch process of their M3 model.

M3 is one of the few native multimodal models in China that has enhanced programming and agent capabilities, and it carries extremely high expectations.

I was initially invited to follow the model's go-live journey, and this extremely flat organization operates with full transparency—so much so that I practically became a ghost wandering through their daily operations.

Yet right after M3's release, MiniMax quickly found itself embroiled in another debate over how to price the model product. A confluence of complex factors collided all at once, plunging the company into a minor crisis.

That's how I ended up experiencing MiniMax's "darkest hour" up close and in real time.

This unique experience and the access I was granted are quite rare. It gave me a window into a rapidly growing, already public model company, showing how its members quarrel, reconcile, reflect, and re-evaluate themselves when facing adversity.

Everything unfolded at 10x speed—chaotic, yet brimming with intense vitality. A company's true character is often "exposed" precisely during these difficult moments.

Now, except for the most sensitive confidential details, I want to present the full story.

1

The Quarrel

The quarrel arrived exactly as anticipated.

It erupted in the 100-person "war room" group chat that MiniMax had set up for the launch of their new M3 model. M3 is a native multimodal model with a 1M-token context window. MiniMax had placed high hopes on it, aiming to push their model capabilities to a whole new level.

Early on the morning of June 1st, M3 officially launched. It was undoubtedly a model with a unique market positioning. During the morning trading session of the Hong Kong stock market that day, MiniMax's share price once surged by more than 7%.

However, the storm struck immediately afterward.

Just half a day later, the group chat—where everyone had stayed up all night excitedly celebrating the model's release—was flooded with a constant stream of external criticisms being shared by team members.

The controversy centered on the pricing scheme that had been rolled out simultaneously with the model.

Alongside the M3 launch, a new billing system went into effect: the previous subscription-based Coding Plan was switched to a new Token Plan that charges per token. Due to the model's upgrade, many users soon discovered that for the same level of usage, their credits were being consumed far faster than expected. More critically, this adjustment was made without any explanation: no SMS notifications to users, no in-app messages, and even the explanatory information on the official website was unclear. Many individual developers only realized the rules had changed after logging back into their accounts.

Dissatisfaction began to ferment. Some users rushed to complaint platforms to demand refunds, others announced they would not renew their subscriptions, and they vented their frustration across social media.

By that point, I had already "infiltrated" MiniMax for a short while. I watched these screenshots flood into every internal group chat almost instantly.

A group of people who had barely slept quickly gathered to discuss how to explain the situation to users. After aligning on the design rationale internally, they immediately realized they had failed to communicate properly beforehand. Solutions were proposed rapidly, and that same evening, an apology announcement was released—acknowledging that the team had not fully communicated with users before the adjustment, that the handling of weekly quotas for long-time users was improper, and admitting "we did not do our job well enough."

But even so, MiniMax's stock price continued to decline throughout the day, closing down 15.71%.

Everyone was working nonstop, and everyone could feel the mood shifting.

The day after the model's launch, anxiety and frustration reached a breaking point. A technical lead finally directly questioned colleagues from the Open Platform team in the group chat, pressing them on how the pricing had been designed in the first place—this department serves model product users and enterprise clients.

Naturally, the other side pushed back, and the argument broke out.

Beneath this tense standoff lay a massive gap in expectations. Over the past several months, the entire MiniMax team had poured all their energy into this make-or-break model, but now all attention had veered off course. Everyone just wanted to figure out what had gone wrong and fix it as quickly as possible.

Having witnessed the team's high hopes before the launch, this looked to me like the beginning of a cycle of mutual blame.

Quarrels are unavoidable in any corporate organization, and I've always seen them as cracks that offer a clearer view into a company's true character. Now, it was unfolding right in front of me—unexpectedly, yet perfectly naturally.

They went back and forth fiercely over how to tier the pricing, how to map the usage quotas, and whether the design process had ever truly considered the user's perspective. Yet within an hour of this intense emotional confrontation, the package design had been fully updated.

It was a highly efficient quarrel.

I later discovered that such disputes are not uncommon at MiniMax. They happen in large group chats, in meeting rooms, and even in top-level strategy discussions—more often than not, playing out openly in front of anyone who happens to be there.

After the situation calmed down somewhat, I arranged to meet the two main people involved in the argument in a conference room.

I was talking to one of them when the other walked in. I assumed there would be some awkwardness between them, but there was none at all.

"I knew he was right back then, and in fact, we ended up making the changes he suggested," one of them said, pointing at the other. "But I still couldn't help but argue with him in the moment."

"At the end of the day, there's no bad blood between us after a fight," the other added. "We just bicker over extremely granular internal parameters."

I later learned that on a separate occasion, when the team revisited the original pricing controversy, that same technical lead directly defended his Open Platform colleague—who was not present at the time—insisting the overall direction of the pricing scheme was sound. The real problem, he argued, was that the colleague had only joined the company two or three months earlier, so he had no experience with the earlier versions of the packages or the previous user sentiment.

"What he lacked was context. That's everyone's responsibility."

2

Context

At MiniMax, most people trickle into the office around 10:30 AM. Lunch and dinner are ordered by the company, and there's no clock-in requirement—except on weekends, when employees do need to sign in for overtime. This reversed attendance policy is designed to protect employees' personal time.

By noon, the workspace gradually comes alive. The three-floor office in Caohejing is laid out like an open-plan dormitory, with meeting rooms named after various stars scattered throughout the space—they're almost always fully booked, so employees often struggle to find an empty room. Several larger meeting rooms are used for receiving guests and for the weekly all-hands meeting. This weekly Friday lunchtime gathering invites industry professionals to share their insights. Most recently, a professor from Shanghai Jiao Tong University who had returned from a research trip in the US talked about the relationship between psychology and AI. Employees submitted questions online, discussing "Anthropic's observation that models exhibit clear anxiety and neurotic traits—how can this manifestation be explained psychologically, and how can we trace it back to training methods and data?"

The last workday before the model's launch was also a Friday, and the all-hands meeting proceeded as usual. The guest that week was the screenwriter of *Ren Fan Xiu Xian Zhuan*.

Many people at the company are huge fans of the series, including IO (the internal nickname for MiniMax founder Yanjun Yan, which we'll use throughout this article). The most popular question in the online Q&A section had just two words: "Release more!"

The atmosphere was relaxed and cheerful. 48 hours later, it would be MiniMax itself that everyone was urging to "release more."

On June 1st, the eve of M3's launch, most MiniMax employees gathered at the Shanghai headquarters. That night, some team members monitored service stability in the workspace, others tracked checkpoints in the group chat, and a few huddled around a table of crayfish in a small meeting room, staying up late to finalize the last details.

Even so, the company's most natural way of collaborating remains online. Key group chats are not limited by department or business function—anyone who has the context for a task is added to the corresponding group. For important matters, meetings are called immediately, starting quickly and ending just as fast. Critical information often pops up in group chats all at once, visible to every relevant person simultaneously.

Context is the core value that this company implicitly emphasizes in every action. To that end, it strongly encourages the free flow of information.

My first impression of this organizational style was chaotic yet vibrant: countless group chats, rapid information flow, and very few buffers between people. It feels like an open public square.

In the early hours of June 1st, as the new MiniMax M3 model entered its final pre-launch phase, I watched all the algorithm and engineering team members join a single shared document, updating the benchmark scores from the latest checkpoint and adding the final technical descriptions of the model's inner workings.

Watching dozens of cursors blinking, moving, and editing simultaneously across the screen felt like the entire company was squeezed onto a single sheet of paper.

IO was right there in the document too.

During those pre-launch wee hours, I saw IO added to a communication thread where algorithm and key technical staff were finalizing core details. The discussion was lively, and everyone was voicing their opinions.

IO mostly just listened—this wasn't a dynamic where everyone took orders from him. But in the end, once enough context had been shared, he would speak up and deliver the final decision.

After observing and experiencing the company's collaborative style extensively, I realized this is standard practice—

When someone's input is needed for an important task, they're called in immediately. The team trusts that with full context, anyone can quickly contribute and find new information, and IO is no exception. But once everyone has all the context, he remains the one to make the final judgments, decisions, and trade-offs for the most complex and critical matters.

During the intense 48 hours before the model's launch, I could directly feel the excitement rippling through the group chats.

M3 was no ordinary update. It embodied so many of the company's aspirations. That's precisely why, when their excitement and high expectations were met with doubts, attacks, and even abuse stemming from the "lack of context" surrounding the token plan, it felt like a "darkest hour" for many on the team.

3

The Root Cause

Back in March, OpenClaw sparked a widespread wave of Agent enlightenment. MiniMax's then-current model M2.5 seized the moment when everyone was frantically "raising shrimp" (running Agent applications): a model with exceptional cost-effectiveness and strong capabilities was the perfect match for these products, and MiniMax achieved a highlight moment in its model development journey.

But the model landscape shifted rapidly. On March 18th, MiniMax launched M2.7, a model with an activated parameter count of only around 10B.

In the following two months, Kimi open-sourced K2.6, DeepSeek's V4 delivered a 1M-token context window, and Zhipu's GLM-5.1 announced that its programming capabilities were approaching Claude Opus 4.6. Across the ocean, Anthropic had released Claude 5, and OpenAI's GPT-5.5 was also on the table.

M2.7's users began to feel that the model's positioning was no longer sufficient for their needs.

A product line employee shared backend user feedback with me: 60% of the issue categories at that time pointed to model performance limitations.

One thing that outsiders often overlook is that MiniMax was among the earliest companies in China to invest in large model training, and it has consistently allocated the most resources to model development. However, because MiniMax also builds its own end-user products, most people formed their first impression of the company through those products, rather than the models themselves.

Once models crossed a certain intelligence threshold, MiniMax M2 became the first model that the public recognized for its impressive intelligence level. Based on the company's actual internal KPIs and priorities, enhancing model intelligence was the top objective, with all resources directed toward delivering a more powerful model.

M3 was a critical milestone in this long-term goal.

"All our attention was focused on the model's intelligence itself," multiple employees from different business lines described the company's collective mindset in the months leading up to the launch.

As the controversy over the token plan continued to escalate over the course of a single day, I watched the company release a model they had waited months for—initially earning some admiration from the tech circle—only to trigger user anger over the billing plan, issue an apology, and ultimately lose double-digit percentage points in market value.

In the days that followed, feedback shifted from isolated messages to a constant stream of complaints flooding every group chat. Follow-up solutions were rolled out one after another, but many criticisms began to stray far beyond the billing plan itself. These criticisms turned into outright abuse, and the team started noticing bot-like accounts with coordinated behavioral patterns acting as nodes for emotional viral spread.

All of this was deeply confusing. Most importantly for MiniMax, these discussions were distracting people's attention from the model's own technical innovations and new capabilities.

At the same time, they realized in the most painful way possible how devastating the consequences of a lack of advance communication and missing user context could be.

It might seem incomprehensible, but to a large extent, these issues stemmed from inexperience, and from the company's overly idealistic approach to action.

As mentioned earlier, M3 was MiniMax's starting point in pursuit of a far more powerful model. But in previous, lighter-weight model launches, the team had developed a habit: the most important thing was to get the model into everyone's hands as quickly as possible, release it right after training finished, and prioritize user feedback on the model's intelligence level.

After speaking with multiple employees involved in the decision-making process, I discovered that the design and change to the billing system had actually been in the works for a long time. It was a byproduct of the company's pursuit of better model capabilities, not a deliberate commercialization move as it might appear.

Back in March or April, as the model in training kept getting stronger, the team realized the billing system for the new model needed an overhaul: the old system was too complex, requiring separate usage quotas for every new model added, and there was another layer of complicated conversion calculations between usage counts and subscription packages. Platform development employees often faced this problem—users would point at the long list of configurations and ask, "What is this? I have no idea what this means."

On the "endgame" of model development, the internal consensus is that using high-quality models will become as common as accessing utilities like water, electricity, and gas, with tokens becoming the service itself.

Since the current subscription plan clearly wasn't the future, it had to change. But how quickly should that change happen?

MiniMax's underlying decision-making logic became clear once again: if this is inevitable sooner or later, why not change earlier?

"If we didn't make the change, users might not even be able to use our new model properly," an employee who participated in the discussions and decision-making recalled.

This isn't a dilemma unique to MiniMax. In fact, the entire industry was re-pricing in the first half of 2026. "When an industry is growing at 20x speed in six months, historical legacy issues are unavoidable. You can't stop to perfect every single detail before moving forward."

All of this became even more complicated for MiniMax.

The company is only four years old, and everything about it is evolving rapidly. But as I observed up close, it still operates like a startup internally.

Yet it's also the fastest large model company to go public, facing scrutiny from the complex capital market and commercialization pressures. Problems that most startups would ignore get amplified several times over, bursting forth unexpectedly all at once.

Everything that happened in June was a concentrated manifestation of this complexity—a painful growing pain brought on by model advancement, crashing