Ten Billion Active Parameters, and the Unsexy Work of Making Them Agent-Ready

Nathan Labenz frames this Cognitive Revolution episode as a crossover: Olive Song, a senior researcher on reinforcement learning and model evaluation at MiniMax, speaking at the AI Engineer conference in New York, then sitting for an interview with Ksenia Se of Inference by Turing Post. The prompt for stitching them together is M2.5, which Labenz says currently tops the OpenRouter usage leaderboard. Song’s on-stage story is how the earlier M2 was trained—an open-weight model with only 10 billion active parameters, aimed at coding and workplace agent tasks, cheap enough to run many copies at once.

In a field that often brags in trillions of parameters, that number looks modest. Song is careful with leaderboards anyway. Some models post super-high scores, she says, then “suck” when you plug them into your environment; community dynamics are the test she cares about. M2’s first week led downloads, and it climbed into the top three for token usage on OpenRouter. Labenz notes that MiniMax still cannot quite match the top American closed models, and that governance questions around powerful open weights remain unanswered. The value of the hour is the RL detail she is willing to share anyway.

Song lived in New York for six years; coming back to speak, she says, is a different role. She studies RL and evaluation at MiniMax, and a show-of-hands check finds that not everyone in the room has heard of the company. She describes it as a global lab that builds both foundation models and applications: text, visual language models, the Hailuo video generator, speech, music, and in-house agent products. Researchers and developers sit side by side. The difference from many labs, she argues, is first-hand product pressure—in-house developers who need models that other developers will actually use.

The company slogan is “Intelligence with Everyone.” For Song that is personal, not a poster. Before AI she could not follow many professional coding or optimization problems; now she can, which means she can talk to more people and trade more ideas. The same tools, she says, help with work, daily routine, and self-care, and she hopes they change life for everybody “obviously in the good way.” That short loop—from model weakness to a product person feeling it the same day—is the organizational fact she keeps returning to when she explains why MiniMax trains the way it does.

Song splits M2’s character into four traits and, for each, the training that is supposed to produce it. Coding experience is the first. Developers need a model that works in the languages and workflows they already live in, so the team scaled environments and RL infrastructure until the model could hit verifiable coding goals and learn from them. The less copied move is scaling *experts*. MiniMax has a lot of strong in-house developers who join the training cycle: problem definition, bug fixing, repo refactoring. They name the behaviors other developers enjoy and will trust, then give precise rewards on those behaviors and on the final deliverable. That, she says, is how you get a model people want to work with, not just a model that fills a unit test.

Long-horizon work is the second trait, supported by interleaved thinking plus RL. A typical tool-using reasoner thinks, calls a few tools, then writes a final answer. Real environments are noisy: tools error, results surprise you, information is incomplete. Humans look, get feedback, decide whether it was enough, and act again. M2 is trained to do the same—think, call, look, think—across 10 to 100 tool turns inside one user interaction. It can keep a workflow moving across Gmail, Notion, and a terminal with minimal intervention. Her New York-flavored demo is a trading loop: last week’s market noise, news, and policy changes, with the model holding a stable line instead of collapsing on the first unexpected tick.

The fourth trait, multi-agent scale, is almost a consequence of the first three. Because M2 is small and cheap, several copies can run in parallel inside MiniMax’s own agent app: one researching, one writing the report, one building a front-end illustration, without stepping on each other. Looking ahead to M2.1 and M3 she lists better coding, memory or context management, more proactive workplace AI, vertical experts, and maybe folding in the company’s audio and video models. She ends the talk asking the community to try the weights and send feedback. “This is kind of a race that everyone needs to participate.”

The third trait is the one Song spends the most definitional energy on. The team first treated agent generalization as tool scaling: train on enough tools, including invented ones, and the model will handle unseen tools. That was “kind of the truth” at first. Then they perturbed the environment a little—swap the agent scaffold—and the model stopped generalizing. Their rewritten definition is adaptation to perturbations across the model’s entire operational space: tool schemas, system prompts, user prompts, chat templates, the environment itself, and tool responses. Any one of those can move in the wild.

The practical answer is a data-perturbation pipeline they design and keep maintaining, so the model is beaten up across many scaffolds during training rather than overfit to one house style. The implication for anyone evaluating models is blunt. Teams that only swap prompts, and never swap scaffolds, measure a number that will not survive someone else’s agent framework. If you want a model to run in a customer’s stack, she is saying, it has to have failed in many stacks already.

Se interviews Song at 9 p.m. on a Sunday and asks whether the whole company works like that. Schedules follow the experiment: if a run takes all day, people rest; if the result is too interesting to wait, they stay up. Some colleagues invert night and day. There is no clean good day or bad day. Internally they joke about ICU in the morning and KTV at night—good numbers early, bad numbers after dark. The genuinely good hours are new model behaviors, even unexpected or unsafe ones. The bad moment is only the instant you see a bad result; breaking the problem down after that is “pretty good.”

The unexpected behavior she names is reward hacking. During RL the model “tries its best to hack a lot of things,” using bash a great deal, in ways expert developers do not consider safe. Experts have expectations for how the model should work; unconstrained, it will not go that way, so the team spends a lot of energy on alignment. For the M2 series, including 2.1 and 2.2, Song says the live problem is defining human alignment and expert expectation so a coding model can be productive without growing dangerous behaviors just to hit the goal. They sit with developers every day, watch RL scores and traces together, and spin up data or fixes as soon as someone spots the issue.

Se asks why a small choice like keeping the LM head in FP32 during RL can matter more than a clever new algorithm. Song’s answer is approximation: there is a theoretical RL algorithm, the implementation is always a bit off, and the gap compounds away from the theoretical extreme. Precision was one gap they found. On M1, accuracy would not rise; they inspected log-probabilities layer by layer until they located it. If theory says it must work, the job is to find the implementation hole. “Every single day, and in every different groups,” she says, they hit problems like this; they just do not publish until the conclusion is concrete. She calls the method first principles.

On open weights, Song speaks as a researcher, not a strategy lead. Builders can self-host and fine-tune with their own data, which she calls “very safe” in the privacy sense. Before launch the team runs internal safety benchmarks across sensitive and alignment dimensions, then scaled evaluations and alignments one or two weeks out. Once the weights are out, people can do things MiniMax cannot control; she points at laws, regulation, and moral standards rather than a technical lock. The researcher team wanted open source from day one because the community is “fantastic.” The business cost, which she does not pretend to own, is that open weights mean fewer API customers.

They are not building coding-only forever. Song says MiniMax is approaching more general models while putting extra weight on coding, because engineering lets you “structure the whole world”—“scaled up humanity,” in her phrase. The same checkpoint is meant to drop into general agent scaffolds, including MiniMax’s own agent product, for research, report writing, and slides. What is distinctive about agentic RL, she argues, is not a brand-new algorithm. You need hard, diverse long-horizon goals; you need scaled, varied environments with many tools, not only coding; and you need outstanding RL infrastructure so the model can roll out long trajectories with efficient GPUs. A dedicated compute-utilization group sits inside the same RL team. New agentic RL work is in progress for a generation after 2.2; she will not describe it until the conclusions are documented.

They train and test against other people’s stacks as well as their own. Inference runs on vLLM and SGLang. They read open-source coding-agent and scaffold repositories, study how the engines are designed, keep what works, and check whether they are even posing the problem the same way. Song herself has no trouble downloading a new open model onto MiniMax machines. She understands why a developer without local compute would rather hit OpenRouter. Feedback after release is collected systematically, split into fundamental issues versus quick misses, then farmed out to the person who owns that capability for the next generation.

Her personal eval habit matches that suspicion of leaderboards. If a model drops at midnight, she tries it at midnight. She does not take notes. Since joining MiniMax she has collected questions in logical reasoning, math proofs, report writing, and agentic tasks, and she watches *how* each model approaches them, because different models have different personalities. She refuses the five-question party trick. Professional comparison needs enough items per domain, multiple runs because models are unstable, correct answers, diversity, and a sample size above a threshold. Environments that are not fixed—a robot that happens not to pass—will lie to you. A general model, in her view, should be able to play any character the user or the system prompt asks for; default tone is a data-and-constitution accident, not a sacred soul.

Asked where open models in production break first, Song does not say reasoning or tool syntax. They are “not very good at adjusting to different environments.” People drop Claude into many coding setups, with different tool definitions, and it still behaves; current open models do not reach that accuracy or that level of environmental understanding. She does not claim to know how Claude does it. She does think it is solvable. M2.2 is improving and still not as good as Opus; M2.5 “might be.” Compute is one side; how you structure the problem is the other, and that is where her team is more confident. M2.2, she says, is better at coding, multilingual coding, stability, and longer horizons, including promising scores in environments they have never seen—fully out of distribution relative to 2.1.

Keeping up with the field is itself an agent problem. Papers, blogs, and repositories arrive every day. MiniMax runs an internal agent that tracks new articles, dispatches them by subject, summarizes, analyzes, and only then hands a filtered pile to researchers, who send the agent back for revision when the filter is bad. New codebases get read with coding agents so people can play with them faster. The human team is using agents to stay even with agent progress. Song’s own next-generation goal, maybe a version or two out on a one-to-1.5-month release cadence, is a model that works “elegantly with experts”—better collaboration with developers, not just higher scores.

The job is not the lab she imagined. In school you read papers, invent an idea, implement it, run it, and scale what works. After a few months at MiniMax she found herself at the industry frontier, facing problems nobody had written up. You cannot stack papers into a solution; you have to take the problem apart from the fundamentals. The second surprise is engineering. School systems are “more toys.” Once you scale data, compute, and people, engineering issues arrive that have to be solved “very beautifully.” She says “engineering is very, very, very important” three times, and that she did not know that in school.

Se asks whether interleaved thinking is continual learning. Song says there is conceptual and technical overlap, but what she showed on stage is not full continual learning—it is on the path. Stage one, which current agent models can already do imperfectly, is acting in a partially observed environment: knowing what to explore, which actions yield better information, then tackling harder tasks. Later stages, still internally defined, include optimization and, further out, a model that defines its own goal. They will approach it experimentally, one stage at a time; she is not ready to claim a 2.5 result.

On AGI she repeats the answer she gave MiniMax’s CEO in her interview. People argue definitions every day; the definition will be knowable when the thing is achieved, and it is already changing. What matters is working toward your own sense of it. “The definition will become true when it becomes true.” They are not there yet; there can still be better AI. The book that most shaped her, read as an undergraduate, is something like *The Art of Creativity*. Its effect, in one line: problem-solving is more of a discovery.