AI/ES/25/11 Code/LEAD/20 Thu/1220 minimax M2 Olive Song
12:20pm - 12:39pm Talk: Minimax M2
- Person/Olive Song Twitter/X - Senior Researcher, MiniMaxAI
Talk: Minimax M2
- Introducing Minimax's latest AI model and its applications in code generation.
Slide — MiniMax: Global Leading Independent Model Lab + App Developer
- Top banner: Bold claim of being a leading independent model lab + app developer.
- Research row (left→right visual flow):
- MiniMax M2 (red card): “Text Model,” “Efficient Model for the Agentic Era.” Positioned as flagship.
- Hailuo Video (dark photo card): “Breathtaking motion, lifelike emotion.” High-fidelity video generation.
- Speech 2.6 (yellow card): “Real-time response,” “Intelligent parsing,” “Fluent LoRA voice.” Indicates low latency + controllable voices.
- Music 2.0 (purple card): “Versatile, agile vocals,” “Precise instrument control,” “Professional.” Signals multi-instrument, multi-voice capability.
- MiniMax MCP Server (gray card): Unified “MCP” endpoint for video, image, speech generation, and voice cloning.
- Left red text block:
- Models: “Excel in all modalities.” Meta-point that MiniMax spans text, video, speech, music, and MCP tooling.
- Product row (AI-native apps, left→right):
- Agent app: “Minimize effort, maximize intelligence.” General assistant.
- Hailuo Video: “Your stories, amplified by AI.” Consumer video creation tool.
- MiniMax Audio: “Create lifelike speech.” Focused speech synthesis editor.
- Video Agent (NEW): “Vibe videoing with zero-touch.” Automated video creation pipeline.
- Talkie Lab: “Create characters you’ve imagined.” Character-driven AI companion platform.
- Left red text block (product side):
- AI-native apps: “150M users globally.” Scale claim.
- Bottom-right: Tagline “Intelligence with Everyone.”
Slide — M2: Performance
- Top text: Benchmarks come from the Artificial Analysis Intelligence Index v3.0 (10-eval suite: MMLU-Pro, GPQA Diamond, Humanity’s Last Exam, LiveCodeBench, SciCode, AIME 2025, IFBench, AA-LCR, Terminal-Bench Hard, τ²-Bench Telecom).
- Visual layout: Eight mini-bar charts in a 2×4 grid, each comparing MiniMax-M2 (red) against open-source models (DeepSeek-V3.2, GLM-4.6, Kimi K2 0905, Gemini 2.5 Pro OSS) and Claude Sonnet 4.5.
- Pattern across charts:
- M2 repeatedly leads or ties for first among open-weight models, often approaching or matching proprietary frontier baselines.
- SWE-bench Verified: M2 ~69.4 tops open-source bands (mid-60s) and is close to Claude 4.5.
- Multi-SWE-Bench: M2 ~36.2 ahead of DeepSeek (~30.6) and GLM (~30.0).
- Terminal-Bench: M2 ~46.3 leading OSS peers (mid-30s to low-40s).
- ArtifactsBench: M2 ~66.8 second to GPT-5(Thinking) at 73.
- τ²-Bench: M2 ~77.2 ahead of all OSS (60–75 range).
- GAIA (text-only): M2 ~75.7 ahead of OSS models (63–71) and behind Claude 4.5 (~76.4).
- BrowseComp: M2 ~44.0 significantly above OSS (14–45).
- FinSearchComp-global: M2 ~65.5 ahead of OSS (26–42) and close to GPT-5(Thinking) (~63.9).
- Takeaway: Slide frames M2 as the strongest open-weight model on coding-centric and agent-centric benchmarks, consistently outperforming all open-source peers.
Slide — Training Shapes M2 Behavior for Developers
- Layout: Two vertical stacks (left = model traits, right = training levers) joined by left-pointing arrows showing causality (training → capability).
- Model Characteristics (top→bottom):
- Coding & Dev Experience
- Long-Horizon State-tracking
- Robust Scaffold Generalization
- Multi-Agent Scalability
- Training Methods (aligned right):
- Scaled Environments & Experts
- Interleaved Thinking & RL
- Data & Eval Perturbations
- Small & Cost-effective
- Interpretation: Each training regime directly produces the listed developer-facing capability—emphasizing agentic behavior, long-context reasoning, code-scaffold reuse, and multi-agent orchestration.
Slide — Scaled Multilingual Environments for Data & RL
- Layout: Three vertical columns labeled in red; each with bullet details beneath.
- Multilingual Coding: JS, TS, HTML, CSS, Python, Java, Go, C++, Kotlin, C, Rust, etc. Signals broad polyglot training corpus.
- 100,000+ Envs: Real GitHub repos, issues, and tests used as RL environments, not synthetic toy tasks.
- High-Concurrency Infrastructure: 5k+ sandboxes per 10s window; tens of thousands running concurrently. Emphasizes massive parallel RL throughput.
Slide — Expert Developers as “Reward Models”
- Left block: A rounded pink rectangle labeled Expert developer with roles listed: frontend, backend, data engineers, etc. Represents humans-in-the-loop.
- Center arrows:
- Right-pointing: Feedback → (experts shaping training signals).
- Left-pointing: Developer-Friendly Model ← (model improved via expert judgments).
- Right stack (top→bottom):
- Problem Definition
- Model Behaviors
- Evaluation & Reward
- Environments
- Meaning: Expert developers supply judgments that determine reward shaping, evaluation criteria, behavior expectations, and environment construction—effectively acting as the “reward model” for training M2.
- break
Slide — Interleaved Thinking (10–100 turns per interaction)
- Layout: Two vertical pipelines: Without (left) vs. With (right); each step shown as a colored block. Shared header: Tools Information, System Prompt, User Prompt.
- Without interleaving: The model alternates thinking → tool call → tool response, repeating this rigid cycle until late in the sequence. Only at the end does it perform Final Thinking → Final Content. Thinking never happens after a tool response until the final stage.
- With interleaving: After every Tool Response, the model immediately re-enters Thinking, optionally produces Content, and only then decides on the next Tool Calling step. This repeats for many micro-turns (10–100) inside one user request.
- Effect: Tool output is always followed by fresh model reasoning, enabling tighter feedback loops, higher-quality plans, and more adaptive multi-step tool use.
Slide — Interleaved Thinking Enables…
- Left column (two outcomes):
- Adaptation to Environment Noise — model can re-evaluate after every perturbed tool result, smoothing volatility.
- Focus on Long-Horizon Tasks — sustained reasoning across many micro-turns without collapsing into a single-shot plan.
- Right visual: An “AI Stock” live-trading dashboard with multiple model traces (M2, Gemini 2.5 Pro, Claude Sonnet 4.5, GPT-5).
- Chart shows TOTAL ACCOUNT VALUE trajectories over days; M2’s line maintains tighter control and fewer cascades during shocks.
- Voting + copy-trading UI suggests benchmarked agentic performance.
- Interpretation: Interleaving lets the model absorb noisy, shifting signals and recalibrate continuously, making it more reliable over long, multi-step environments.
Slide — Perturbation Pipeline
- Core layout: A vertical agent-execution stack on the right; multiple leftward arrows indicating where perturbations can be injected into the pipeline.
- Perturbation injection points (top→bottom):
- Tools Information (purple)
- System Prompt (blue)
- User Prompt (green)
- Tool Response (pink, repeated at multiple layers)
- Execution spine: After each perturbed input, the model runs:
- Thinking → Content → Tool Calling → Tool Response → (repeat)
- Eventually ending with Final Thinking → Final Content.
- Interpretation: MiniMax maintains a structured perturbation framework that deliberately introduces noise or variation across prompts, tool metadata, and tool outputs.
- This forces M2 to become robust across many agent scaffolds, not overfit to a single orchestration pattern.
- By perturbing every layer—from system prompt to tool outputs—they train generalization across the entire agent loop, yielding stable behavior even when the surrounding agent system varies.