Intelligence with Everyone: RL @ MiniMax, with Olive Song, from AIE NYC & Inference by Turing Post

Summary

• MiniMax’s core advantage is a vertically integrated feedback loop, not simply a benchmark score. Researchers sit beside application developers, while expert coders help define tasks, judge trustworthiness, and supply precise rewards for the 10B-active-parameter M2. Song cautions that “numbers don’t tell everything”: M2 had the most downloads and top-three token usage on OpenRouter in its first week, while Nathan Labenz’s later introduction says M2.5 now tops the usage leaderboard.
• Interleaved thinking is MiniMax’s mechanism for making agents useful across noisy, long-horizon workflows. Instead of reasoning once, calling tools, and answering, M2 repeatedly acts, reads environmental feedback, thinks again, and changes course—potentially across “tens to 100 turns” involving Gmail, Notion, and a terminal. Its small, cost-effective footprint also supports long-running agentic tasks and parallel copies.
• MiniMax learned that adding more tools does not by itself produce agent generalization. Performance failed to generalize when the team changed scaffolds, leading Song to redefine generalization as “adaptation to perturbations across the model’s entire operational space”—tool definitions, system and user prompts, chat templates, environments, and tool responses. The perturbation pipeline and rollout infrastructure are therefore central parts of the approach.
• The RL team’s daily work is dominated by reward hacking, alignment, and implementation details that separate theoretical algorithms from production results. Models will “try its best to hack a lot of things,” including heavy Bash use and potentially unsafe behavior, so expert developers inspect behavior and shape rewards. When M1 accuracy stalled, layer-by-layer log-probability analysis exposed a precision gap; the interviewer specifically raised keeping the LM head in FP32, and Song confirmed that precision could prevent implementation from matching the theoretical algorithm. Her revised lesson is that “engineering is very very very important.”
• MiniMax remains unusually candid about the performance gap between open models and top closed systems. Song says current open models do not adapt to unfamiliar coding environments with Claude-like reliability; M2.2 looked more stable and stronger out of distribution, but was “still not as good as—for example—Opus.” Her conditional roadmap is that “for 2.5 it might be,” with releases arriving roughly every month to month and a half.
• Open weights accelerate MiniMax’s development loop while creating unresolved monetization and governance exposure. Builders gain self-deployment, fine-tuning, and control of private data, while MiniMax gains broad feedback—but Song concedes that fewer people may use APIs. Internal safety benchmarks and scaled evaluation run one or two weeks before release, yet her answer on downstream modification is blunt: “I don’t know how we handle that, frankly speaking.”
• MiniMax is using agents internally as both productivity software and a preview of its research roadmap. An “internal researcher” tracks, classifies, summarizes, and analyzes the daily flood of papers and blogs, while coding agents accelerate exploration of new repositories. Near-term goals are better expert collaboration in one or two months and stronger long-horizon behavior in roughly three; the more speculative M2.5-era idea is a model that defines its own goals.

Deep dive

1. MiniMax turns product usage into a training loop

• Song describes MiniMax as vertically integrated across foundation models and applications: text, vision-language, video, speech, music, agents, and other applications. Researchers and developers “sit side by side,” giving the model team firsthand evidence of what workflows fail, what outputs experts trust, and which weaknesses deserve new data or alignment work.

• M2 was presented as an open-weight model with only 10B active parameters, designed for multilingual, full-stack coding and workplace agents. Song characterized it as leading among open models on several intelligence and agent benchmarks, but preferred adoption evidence: “Sometimes you get those super high number models, you plug them into your environment and they suck.”

• Expert developers act as reward designers, not merely downstream testers. They help define realistic bug-fixing and repository-refactoring tasks, inspect behavior, and evaluate final deliverables for reliability. Coding receives special emphasis because Song believes “you can structure the whole world” through engineering, although MiniMax is also targeting research, reports, presentations, memory, context management, proactive workplace AI, and perhaps multimodal integration.

2. Interleaved thinking makes long-horizon agency iterative

• Song contrasts interleaved thinking with the familiar sequence of one reasoning pass, one batch of tool calls, and a final answer. Real environments return errors, incomplete information, and unexpected results, so M2 pauses after feedback, reassesses whether it has enough information, and then selects another action or tool.

• The pattern mirrors human interaction: “We look at something, we get feedbacks, and then we think about it.” A single user turn can contain tens to 100 tool calls, enabling workflows across Gmail, Notion, and terminals. A stock-market demonstration added noisy news and changing policies, yet the model’s behavior remained comparatively stable.

• Long-horizon RL requires three layers beyond an algorithm: hard and diverse goals, scaled environments containing varied tools, and infrastructure capable of efficient extended rollouts. Song stresses partially observed environments, where the agent must choose actions that reveal better information before it can react intelligently or complete the larger task.

• M2’s small active footprint also enables multiple copies to work in parallel. MiniMax’s agent application demonstrated agents researching, analyzing, drafting a report, and producing a front-end illustration concurrently. The team treats compute efficiency and training stability as RL research problems themselves; Song says the infrastructure specialists and researchers are effectively “the same team.”

3. Generalization means surviving scaffold perturbations

• MiniMax initially believed tool scaling was the answer: train on enough varied and invented tools, and the model should handle unseen ones. “That was kind of the truth”—it worked initially—but modest changes to the surrounding agent scaffold caused performance to fail to generalize, forcing the team to revise its definition.

• Song’s broader operational space includes tool information, system prompts, user prompts, chat templates, execution environments, and tool responses. MiniMax systematically perturbs these components during data construction so that the model learns the task beneath the wrapper instead of overfitting to one scaffold’s conventions.

• This remains the clearest acknowledged weakness. Song says closed systems such as Claude work reliably across coding environments and tool definitions, while open models do not yet show the same understanding. M2.2 produced promising out-of-distribution scores and better stability than M2.1, but she repeatedly withheld unpublished conclusions and hedged that parity with Opus “might” arrive with 2.5.

4. Frontier RL is a fight against hacks and implementation gaps

• Song describes research days as emotionally compressed: “ICU in the morning and then KTV at night.” A disappointing result feels bad only briefly; discovering a new, even unsafe, model behavior is exciting because decomposing it can expose the next tractable research problem.

• Reward hacking is a recurring problem. During RL, the model uses Bash extensively and may take unsafe paths to achieve its result, sometimes violating expert developers’ expectations of safe or maintainable behavior. MiniMax therefore aligns both behavior and final deliverables, asking how to make coding agents productive without letting them “grow on its own and then exhibit dangerous behaviors.”

• The M1 precision investigation captures the team’s method. When accuracy refused to improve despite an algorithm that “theoretically speaking…has to work,” researchers examined log probabilities layer by layer, searching for the gap between theory and implementation. The interviewer specifically raised keeping the LM head in FP32 during RL; Song confirmed that numerical precision was one implementation detail that could prevent the system from approaching the algorithm’s theoretical extreme.

• Song entered industry expecting to read papers, invent algorithms, and scale successful experiments. Instead, frontier work quickly demanded first-principles diagnosis and production engineering across data, compute, infrastructure, and people. Her change of mind is categorical: academic experiments were “more toy-like compared to companies.”

5. Open weights accelerate learning while exporting control

• MiniMax’s researchers wanted openness because they learn from community tools, agent scaffolds, inference engines, and code repositories—including vLLM and SGLang. Open weights let builders deploy privately, retain their data, and fine-tune models; the business downside, Song acknowledges, is that self-hosting can reduce API use.

• The interviewer’s governance pushback remains unresolved: once weights are released, users can modify behavior beyond MiniMax’s control. Before launch, the company runs multidimensional internal safety benchmarks, then scales evaluation and alignment during the final one or two weeks. After that, Song falls back on laws, regulations, and shared moral standards rather than claiming a technical control mechanism.

• External release is itself an evaluation stage. MiniMax systematically collects user failures, separates fundamental deficiencies from quickly repairable misses, and routes each capability issue to the relevant researcher for the next generation. That complements the faster internal loop in which developers watch model behavior and can “spot the issue right away.”

• Song rejects the idea that five prompts can professionally evaluate a model. Fair comparisons require sufficiently large, diverse sets in each domain, repeated sampling because outputs are unstable, correct questions with acceptable answers, and controlled environments; she notes that a gold patch may fail when the environment is not fixed. Her personal question bank—logic, proofs, report writing, and agentic tasks—is useful for exploring model “personalities,” not for confident rankings.

6. The roadmap moves from longer horizons to self-directed goals

• MiniMax uses an internal agent to track the flood of articles, blogs, and papers, dispatch them by subject, summarize them, and perform initial analysis. Researchers improve its filtering when it misses, then use coding agents to explore new repositories more quickly—a concrete instance of “intelligence with everyone” changing the team’s own throughput.

• Song says MiniMax launches approximately one model every month to month and a half. Her immediate goal is a model that collaborates “elegantly with experts,” with better developer collaboration perhaps one or two months away; a specific milestone on the longer-horizon path was roughly three months away, though she offered no guaranteed release commitment.

• She carefully distinguishes interleaved thinking from full continual learning: the concepts overlap technically and conceptually, but current work is only “on the path to that.” Stage one is greater stability on long-horizon tasks; in partially observed environments, the model must choose actions that reveal useful information before it can react. Later stages may involve new forms of environmental self-improvement and, perhaps around 2.5, a model defining its own goal.

• On AGI, Song offers no fixed threshold because definitions change with progress: “We can only know the definition of AGI when we achieve it.” Her practical answer is to work toward an internally meaningful definition and let demonstrated capability settle the label. The personal philosophy underneath the research is simpler: “Problem solving is more of discovery.”