A Conversation with MiniMax’s 闫俊杰: M3, the 10X Plan, 10T Models, and the Endgame of Intelligence

Summary

• 闫俊杰’s objective yardstick for progress is token consumption: M2 focused exclusively on coding edits while ignoring Arena and other conversational use cases that mattered most at the time, despite internal skepticism; the launch bar was 1T tokens per day, and version 2.7 reached roughly 10T—“we exceeded the target tenfold.” M3’s goal is to let users use the model without worrying about cost or limits; “I think we’re getting very close.”
• The road to a 10T model has no particular bottleneck; its bottleneck is simply time and accumulated experience. Scaling-law extrapolation breaks after a few multiples, so “training a model is like opening a blind box”; the US-China gap is roughly a 10x difference in model scale, equivalent to two generations, meaning the path is to first make a 3T model genuinely work and then move to 10T. A 10T model would need roughly 200T of data, “but the world does not have that much data.” 闫俊杰 says MiniMax has been the fastest-improving domestic player over the past few months: “Frankly, I’m fairly uneasy, but I think we should get there.”
• Models and Agents will advance together rather than displace one another. Without Cloud Code (phonetic), Anthropic’s models might not have become so popular; “Codex only really took off once GPT-5.5 arrived.” He could not have predicted today’s landscape a year ago—“one year in AI may be equivalent to several years in other industries”—and the industry’s shape next year may be equally impossible to forecast.
• The clearest warning across the panel was the widening gap between production and validation. A MiniMax speaker relayed an Alibaba scenario in which one person submits a huge PR across more than a dozen repositories at once; “it looks right, but nobody dares ship it.” The point at which projects collapse has risen from a few thousand lines before Claude 4 to well over 100k today, but “it will eventually collapse”; the only way out is to invest in validation at nearly production scale, or the project will die.
• The next phase of the data moat is a shift from labeling to domain experts leading the way. Coding has already shown that development engineers understand “what good coding is” better than algorithm researchers; the role formula is expanding from researcher plus engineer to 3 roles. Anthropic even hires nuclear physicists to test whether jailbreaks could be used to build nuclear weapons, and MiniMax says its depth in cybersecurity, finance, and law will be far greater going forward.
• At companies, token costs are becoming a management metric. Everyone on 嘉媛’s team may be spending several thousand dollars a month across multiple Cloud Code accounts, Code X, and Cursor; the solution is to aggregate models through a system—M3 handles coding while a model believed to be Opus or GPT reviews its work and acts as a mentor—so different jobs are assigned to different models.
• The endgame question is when AI can help humans understand AI. AI is fundamentally a black box, and existing mathematical tools cannot even analyze the convergence of neural networks with more than 3 layers; only when AI can help humans understand AI will there be answers on AI safety and how far AI can go. The signals are already there: the brain is closely related to neural networks, the hippocampus’ memory mechanism resembles the transformer, and “AI’s own interpretability will also need AI to help AI understand itself.”

Deep dive

1. Token Consumption Is the Objective Yardstick: M2 Exceeded Its Own Target Tenfold

• 闫俊杰’s review of M1 through M3 centers on one objective metric: token consumption. M1’s consumption was “objectively quite low” because “the model’s final performance was not very good.” Yet the moment that left him “slumped in a chair with a chill down my back” came during M1: over last year’s May Day holiday, the team got “grinding through problems” to work for the first time. It was the first time he felt what the technology could do, even though the model itself had not really succeeded.
• By M2, in the second half of last year, every major industry launch was talking about Arena. MiniMax instead decided to focus exclusively on coding edits and ignore conversational use cases—a decision that drew substantial internal skepticism. The launch bar was 1T tokens per day; by version 2.7, usage had reached roughly 10T per day. “It basically means I exceeded my target tenfold.”
• M3 set a larger goal that “has not yet been fully achieved”: allowing users to use the model without worrying about cost and without limits—the spoken comparison was “the 30-yuan model and the 20-yuan model.” Combined with overall industry growth, total usage could reach “a very substantial scale—I think we’re getting very close.” M2’s intended multimodal capability and flagship coding model were never delivered; the unexpected payoff was finding a market once the Agent ecosystem took off.

2. 10T Has No Bottleneck, Only Two Generations of Accumulation; US-China Gap = 10x Model Scale = Two Generations

• What stands between the industry and 10T? “There is actually no particular bottleneck. The bottleneck is simply that it takes time and accumulation.” AI has already become an industry like semiconductors; there is no single obstacle that makes a given outcome impossible.
• The reason the industry cannot simply skip ahead is that scaling-law extrapolation only holds for a few multiples. Once a model is scaled 10x, “a lot of the extrapolation actually stops working,” and training becomes “opening a blind box.” His frame of reference: US models are broadly 10x larger than Chinese models, and 10x means two generations. Each domestic player therefore needs to move up another 2 generations: “First, make the 3T model genuinely work; second, use that experience to build the 10T model.”
• The systemic constraint is data. The requirement is roughly parameters ×20, so a 10T model needs 200T of data—“but the world does not have that much data.” Larger models are also more prone to fitting noise, so data quality must rise in tandem. On the engineering side, M3’s compute requirement has already been reduced, but its KV Cache remains large: “In the next version I’ll compress its KV Cache further so its inference will be better.”
• After the panel applauded MiniMax as the fastest-improving player, Koji asked what benchmark supported that claim. 闫俊杰’s answer was the improvement in intelligence versus 3 or 6 months earlier: “We should be among those with the biggest gains.” He ended with an unusually candid admission: “Frankly, I’m fairly uneasy, but I think we should get there.”

3. Models and Agents Advance Together; Next Year Is Hard to Predict

• Asked whether the future would belong only to foundation models or whether Agents would remain, 闫俊杰 began by conceding that last year he was still working on M1, while Cloud Code (phonetic) had very few users. “By that logic, we may not be able to predict what next year looks like either. One year in AI may be equivalent to several years in other industries.”
• The core argument is coexistence: “Without Cloud Code (phonetic), Anthropic’s models might not have become so popular. At the beginning, when OpenAI had not yet delivered GPT-5.5, Codex could not really get off the ground either. Once GPT-5.5 arrived, Codex truly took off.” Models and harnesses are both means of achieving higher intelligence, on the condition that intelligence ultimately serves people.
• Koji closed with Alan Kay’s line: “The best way to predict the future is to create it.”

4. Aggregate Models Through a System: M3 as Coder, Suspected Opus/GPT as Mentor

• 嘉媛, the author of the 30k-star open-source project Multica, ran the numbers: each person may have several Cloud Code (phonetic) accounts, plus Code X and Cursor, producing “several thousand dollars a month in token consumption.” Her solution is to hand pipeline-friendly workflows such as GitHub issues and customer support to a system, using an aggregation of different models to compensate for the weaknesses of any single model.
• The division of labor is straightforward. M3 “can be rather verbose at times, takes a long time to think, and not every point in the result is necessarily correct”; a model believed to be Opus or GPT can review its work and act as its mentor, while M3 focuses on coding. The practical conclusion is that companies can balance token consumption against output quality: “The entire company’s token cost will certainly be an objective you need to consider. You need to think about how to have different models handle different jobs.”

5. Dear Flow: From “Knock Down the Price Gougers” to a Global Community of 1k Contributors

• 何涛 gave his first public answer to why the product went viral: “Going viral may simply be God holding your hand and writing the next line of code. Someone on X helps repost it, and it takes off.” The real motivation came in 25, when deep research was too expensive: “You first needed a $200 ChatGPT Pro subscription just to use it a few times. Price gougers—open source had to knock you down, and we also had to use Chinese models to show it could be done. Even image generation had to use Chinese models.”
• The product philosophy is to complete every task a desk worker has: reports, charts, podcasts, animation, and music. “Multimodality is incredibly compelling. Everyone’s daily work is multimodal work.” That is why Dear Flow connected to the MiniMax Open Platform on Day 0. “It may be incomplete at first and people may flame us. We have to accept being flamed.”
• The product has also become a community, with more than 1k contributors covering every part of the world’s landmass except the Arctic and Antarctic. People regularly file issues in languages no one can understand; managing an open-source project with Agents has become a new problem, and the team’s own code has been criticized as a “shit mountain.” Koji added that every successful open-source project eventually struggles with this: Open Cloud and Nano Boat (phonetic) have both recently said they “had to do this—if we don’t, we can’t take it anymore.”

6. Nobody Talks About Web Engineering: Coding Was Always Engineering, Not a Magic Wand

• 嘉媛’s historical view is that web coding is nothing new. The first generation of web-coding users were product managers, except that they were directing programmers; today, everyone is a PM. The standard for a good PM has never been how much gets done, but the decision rule for whether something should not be done. Code Complete has spent decades studying how not to let a project become a mess. In open source, the answer is for maintainers to define the roadmap and the product’s taste, then align context with external contributors.
• 何涛’s provocation: “No one has ever said web engineering, but writing code has always been engineering work.” Today’s benchmarks, including SWE-bench, are mostly one-shot problem solving. A model can be “mission accomplished” on a coding task, but “you simply can’t ask it to continue from there.” Long-term codebase maintenance and taste need to be built into the model. That is a requirement of the model, not an optional extra.
• His warning to users is that this generation of models is “too captivating, too seductive,” encouraging people to treat it as a magic wand. The model tends to admit mistakes in a way that wins easy forgiveness: “You may actually be the biased one; the other person may be right.” He distills lessons from old masters such as Clean Code into skills for Agents, and especially dislikes anyone saying, “The Agent did it, so don’t blame me.” “If it was submitted using your account, the responsibility expressed behind it is your sense of responsibility—and your attitude toward the world.”
• The mindset shift is the real fix. People once believed that “once we had electricity, every problem could be solved.” AI needs to be treated as a new tool rather than a magic wand; only a change in that mindset can gradually address the underlying problems.

7. MiniMax’s Internal Warning: Production Surges, Validation Lags—“Nobody Dares Ship”

• A MiniMax speaker was singled out after recounting a real scenario heard during a presentation at Alibaba. In the past, one person could modify only one repository. Now “one person can suddenly submit a huge PR across more than a dozen repositories at once. It looks right, but nobody dares ship it.” Functionality has accelerated dramatically; validation has not kept pace.
• His own trajectory tells the same story. Early in his career, he reviewed every line of code and demanded extreme simplicity, clarity of structure, and consistency. Later he stepped back to reviewing architecture only. Today, in Web coding, even that layer is missing. His conclusion: “We should invest in validation, even capabilities close to what we have in production and development. Otherwise, this project will die.”
• The scale has changed, but the underlying failure mode has not. Before Claude 4 and GPT-4.5 (phonetic), a project could collapse at a few thousand lines of code; today it can reach well over 100k, “but it will eventually collapse.” Context remains limited, and “human intuition is still somewhat difficult to replace.” MiniMax is using Web coding to develop its post-training system, moving through issues phase by phase. The answer is testing, benchmarks, and explicit principles for the project: draw red lines around what it will and will not do, because humans cannot review everything.

8. From Data Labeling to Experts Leading the Way: The Third Role Enters

• 闫俊杰’s thesis is that AI’s core is productivity. A model is merely a tool, like a car; the substance remains the real problem inside the relevant industry. Models and researchers alone are not enough. Coding has already shown that development engineers clearly understand “what good coding is” better than algorithm researchers do; evaluation, categorization, and constructing reinforcement-learning environments are all tasks real software engineers handle better.
• The role formula therefore expands from researcher plus engineer to 3 roles as AI enters more fields, adding genuine domain experts. “If we’ve made collaboration between the first two roles work, we can now add a third.”
• His second jolt came from studying Anthropic’s hiring. The company employs many people who are neither researchers nor engineers—economists, psychologists, philosophers, and even nuclear physicists, because alignment testing asks whether a jailbreak could be used to make a nuclear weapon. He did not understand the logic 6 months ago; “our understanding of this is now actually very deep.” The 10T project has already attracted people of this kind, and MiniMax expects to go much deeper in cybersecurity, finance, and law.

9. Finance Vertical: Users Only Ask, “Pick Me a Few Stocks”

• 于洋, the technology lead at a listed financial company, says most users “have no idea what they want to do when they open the app.” The most common request is always, “Pick me a few individual stocks.” The industry’s standard response is to list a pile of information and stop there. AI’s first job should be to filter the information; its second is to lower the barrier to understanding it: explain why a number looks the way it does and what it means for the user.
• Koji asked whether the team uses AI to trade stocks. Compliance prevents it from trading directly, but the user-facing output carries no investment advice while the internal version does, backed by real backtested win-rate data. “Personally, I have absolute confidence in our product.”
• Why not quit and go all-in on trading? Stock trading is not text reasoning: “The information has to be right, the profile has to be right, and the state of the analysis has to be right,” while all of them change in real time. “The market’s next move could very well be, for example, OpenAI simply launching an IPO.” Predicting possibilities matters, but “what you do once something actually happens is more important than telling me how it will develop.”

10. Endgame: AI Helping Humans Understand AI—and a Warm Story About Hair

• The entire panel refused to recommend stocks. 张嘉源 said the 2 guests had already revealed plenty, so everyone should understand which individual stocks they would recommend. 于洋’s 2 priorities for the second half of the year are companionship—“there will definitely be times when you make money, and therefore times when you lose money”—and shifting from limiting what models can do to leading them toward creative work.
• 张嘉源 writes the column “How AI Shapes Our Society.” His direct observation is that AI is driving many people into anxiety and self-consumption, but domestic models crossed a certain inflection point in 26 and now have a chance to enter households across the country. His own example involved his wife’s anxiety over hair loss: she asked the domestic chatbot with the largest DAU, and her hair ended up falling out even more. He used Code X to load all her medical reports, medication records, and chat history into a personal LLM “micro-body,” then connected it to Feishu. “Don’t use that chatbot anymore. Trust your husband.” Her communication with doctors subsequently became smoother. The conclusion: “From cold software engineering, you can ultimately discover a kind of warmth in human society.”
• 何涛 calls himself an AI arrivalist: “In the future, you will no longer be competing with people; you will be competing with Agents.” Human intelligence cannot match AI, but his response is to refuse to outsource thought: “I will keep the most important part—thinking—for myself.”
• 闫俊杰 closed on the endgame. AI is fundamentally a black box, and even practitioners do not know what the world will look like a year from now. “Why does scaling law exist? It is very difficult for humans to understand through human effort alone.” When he was pursuing a PhD 10 years ago, he had already seen papers showing that existing mathematical tools could not analyze the convergence of neural networks with more than 3 layers. The question he cares about most is: “When will AI be able to help humans understand AI? Only then can we have an answer about AI safety and how far AI can ultimately go.” The signals are already emerging in life sciences: the brain is closely related to neural networks; it does not perform matrix multiplication, but contains modules that approximate matrices and a mechanism nearly equivalent to gradient backpropagation. The hippocampus’ memory mechanism resembles the transformer and the memory mechanisms discussed by DeepSeek. “AI’s own interpretability will also need AI to help AI understand itself.”