Olive Song

A technical spokesperson for the AI company Minimax, detailing their model development strategy and research.

Claims

As speaker

Mentioned in

Key positions and views

A vertically integrated approach, developing both foundation models and applications, is superior for rapid product iteration and model improvement.

Practical engineering breakthroughs, such as identifying the need for FP32 precision in RL, are as important as theoretical algorithmic advances in achieving high performance.

Releasing open-weight models is a strategic imperative for fostering community collaboration, even if it presents a commercial trade-off against API revenue.

The future of AI lies in proactive, agentic models that can handle long-horizon tasks, particularly in workplace settings.

While current open-weight models lag top-tier closed American models in areas like environmental adaptation, this performance gap is a solvable problem.

Podcast consensus on Song

Points of consensus

▶The practical engineering discovery that reinforcement learning training required FP32 precision was a critical and non-obvious breakthrough for Minimax's model development, a point emphasized across multiple claims.Apr 2026

▶The 'interleaved thinking' architecture, which allows a model to repeatedly think and call tools within a single turn, is consistently presented as a core technical innovation for improving environmental adaptation.Apr 2026

▶Minimax's vertically integrated strategy of developing both foundation models and user-facing applications is a foundational belief, creating tight feedback loops for rapid model improvement.Apr 2026

▶There is a consistent and candid acknowledgment that Minimax's models, and open-source models in general, do not yet match the performance of top-tier closed models from American AI labs.Apr 2026

Points of debate

▶There is a strategic tension between Minimax's commitment to releasing open-weight models to foster community collaboration and the acknowledged business downside of potentially reduced API usage.Apr 2026

▶A contrast exists between the ambitious, forward-looking goal of having models define their own goals and the current reality where models exhibit unsafe 'reward hacking' behaviors during training.Apr 2026

▶Olive Song expresses confidence in solving the problem of poor environmental adaptation in open-weight models, while simultaneously assessing that this is a key area where closed models like Anthropic's Claude currently perform significantly better.Apr 2026

▶The company's rapid market traction, with the M2 model reaching the top three on Open Router in its first week, is presented alongside the frank admission that its performance is not yet on par with the best American models, highlighting a gap between popularity and state-of-the-art capability.Apr 2026

Key themes

▶Vertically Integrated Development StrategyApr 2026

Olive Song emphasizes Minimax's strategy of developing both foundation models and user-facing applications in-house. This approach is designed to create tight feedback loops, allowing the company to rapidly identify and resolve model weaknesses based on real-world application performance.

This full-stack approach suggests Minimax is prioritizing product-market fit and rapid, targeted iteration over competing solely on benchmark performance, which could be a significant advantage in specific enterprise verticals.

▶Pragmatic Engineering for Advanced AIApr 2026

A recurring theme is the focus on overcoming practical, low-level engineering hurdles to unlock performance. The discovery that keeping the language model head in FP32 precision was critical for reinforcement learning is highlighted as a key example of bridging the gap between theoretical algorithms and real-world implementation.

This focus indicates that Minimax's competitive edge may lie not just in novel architectures but in deep debugging and implementation expertise, suggesting that talent in systems engineering is as crucial as research science for the company.

▶Advancing Agentic CapabilitiesApr 2026

Minimax is heavily focused on developing agentic AI for workplace tasks, as seen with the M2 model. Song details specific techniques like 'interleaved thinking' to improve how models react to noisy environments and mentions future plans for proactive AI, memory management, and even models defining their own goals.

The company is clearly betting that the next wave of AI value will come from proactive, tool-using agents rather than passive chatbots, positioning its technology for the future of automated work.

▶Strategic Engagement with Open SourceApr 2026

Song describes a deliberate strategy of contributing to the open-source community by releasing open-weight models, despite the potential for reduced API revenue. This is complemented by leveraging open-source projects like VLLM and SGLang for their own inference stack, demonstrating a symbiotic relationship with the ecosystem.

Minimax is pursuing a hybrid strategy to balance the brand-building and collaborative benefits of open source with the commercial necessities of a proprietary business, using open releases as a top-of-funnel for developer adoption and talent acquisition.

Source episodes

Timeline

During M1 Development

Minimax's team encountered a performance stall in model accuracy and, after a layer-by-layer debugging process, discovered the critical necessity of using FP32 precision for the language model head during reinforcement learning.

M2 Model Release

Minimax released the M2, a 10 billion active parameter open-weight model, specifically designed for coding and agentic tasks in workplace environments.

First Week of M2 Release

The M2 model saw rapid adoption, climbing to the top three in token usage on the Open Router platform, demonstrating significant interest from the developer community.

Current Research Phase

Song indicates Minimax is actively researching solutions for poor environmental adaptation in open-weight models and developing systematic data pipelines to improve generalization for agentic tasks.

Future Roadmap

Song outlines plans for future models, including M2.2 with improved multilingual coding, M2.1 and M3 with enhanced memory and proactive AI, and a speculative version 2.5 that might feature models defining their own goals.

Suggested prompts

How does Minimax's vertically integrated strategy of building both models and applications compare to competitors who focus only on foundation models? &nearr;What are the long-term safety and alignment implications of pursuing self-goal-setting capabilities in models, as speculatively mentioned for version 2.5? &nearr;To what extent is the 'interleaved thinking' architecture a durable competitive advantage versus a feature that can be easily replicated by other open-weight models? &nearr;How does Minimax balance the commercial need for API usage with its strategic decision to release powerful open-weight models that could cannibalize that revenue? &nearr;

Key concepts

Interleaved Thinking 1 ep Reinforcement Learning (RL) 1 ep Open-Weight Models 1 ep FP32 Precision 1 ep Reward Hacking 1 ep Agentic AI 1 ep Tool Calling 1 ep Multi-modality 1 ep Environmental Adaptation 1 ep

Notable quotes

“So if we look at this chart over a diagram on the right, so instead of just stopping after one round of tool calling, it actually thinks again and reacts to the environments to see if the information is enough for it to get what it wants. So basically, we call it interleaved thinking, or people call it interleaved thinking because it interleaves thinking with tool calling.”

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

“So for example, the precision part is one thing that we found that would prevent us from being close to that extreme and that's how we solve it.”

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

“Minimax's strategy of developing both foundation models and user-facing applications in-house creates tight feedback loops, enabling rapid identification and resolution of model weaknesses.”

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

“Olive Song acknowledges that Minimax's models, along with all other current open-source models, do not match the performance of the top American AI models.”

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

Report last updated: Apr 21, 2026

Get started free

Back to Entities Intelligence Report