New impressive coding agent - Free and Open source

Key Concepts

• Minimax M2: An open-weight language model specifically trained for coding and agentic performance.
• Minimax Agent: An agentic system built on top of the Minimax M2 model, featuring different modes for task execution.
• Agentic Performance: The ability of an AI model to act autonomously, plan, use tools, and execute complex tasks.
• Open-Weight Model: A model whose architecture and weights are publicly available for download and use.
• Multi-Agent System: A system where multiple AI agents collaborate to achieve a common goal.
• Reasoning Traces: The step-by-step thought process of an AI model as it solves a problem.
• Tool Use: The capability of an AI model to interact with external tools or APIs to gather information or perform actions.
• Instruction Following: The ability of an AI model to accurately interpret and execute detailed instructions.
• Hallucination: The generation of false or inaccurate information by an AI model.
• Sparse Mixture of Experts (SMoE): A model architecture where only a subset of parameters are activated for any given input, leading to efficiency.
• Active Parameters: The parameters within a model that are actually used during inference.

Minimax M2: A State-of-the-Art Open-Weight Model for Agentic Coding

This video explores the capabilities of Minimax M2, an open-weight language model positioned as the best currently available for coding and high-level agentic performance. The model is state-of-the-art on several key benchmarks and is independently recognized for its strong performance. Its weights are fully open and downloadable from Hugging Face, offering a compelling balance of intelligence, output speed, and performance at a reasonable price. A significant advantage is the ability to try its new capabilities for free.

Minimax Agent: Interface and Modes

The demonstration focuses on the Minimax Agent, an agentic system built upon the M2 model. This system offers two primary modes:

• Lightning Mode: Optimized for fast task execution.
• Pro Mode: Utilizes a multi-agent system for complex development tasks, which is the focus of the video.

Demonstrating Instruction Following and Agentic Coding

The presenter tests the model's instruction-following capabilities with detailed prompts.

1. Professional Landing Page for SaaS:

• Prompt: "Create a professional looking landing page for SAS."
• Execution: The Minimax Agent, in Pro mode, demonstrated impressive results, producing one of the most detailed websites seen from an open-weight model.
• Process:
  • The model first shows reasoning traces, outlining its thought process.
  • It then generates a development plan.
  • It creates a requirements.txt file for itself.
  • A key feature highlighted is that the model tests every step during development, which increases execution time but leads to impressive results.
• Output: A deployed web app with a professional landing page.
  • Specifics: The website included neat animations on button hover and created additional cards.
  • Live Demo: The generated website included a "run code" option for a live demo.
• Limitations:
  • Did not implement the toggle between dark and light mode, although the buttons were present.
  • The arrow animation for expanding/collapsing sections was not fully implemented as expected.

2. Multi-View Workflow Builder:

• Prompt: A complex prompt, structured as a JSON file, to build a multi-view workflow builder that converts natural language instructions into deployable workflows. This is conceptually similar to OpenAI's Agent Builder but driven by natural language.
• Execution: The model generated a detailed plan and proceeded with execution.
• Process:
  • The prompt specifically instructed the model to use the Gemini model and its API.
  • The Minimax Agent typically presents a plan and seeks user permission before execution due to the time involved (15-20 minutes).
  • The model created files, iterated through them, and worked on both the front-end and back-end.
• Output: A functional workflow builder.
  • Interface: Included an input text box for instructions, four expandable nodes (with the potential for more), and a code view.
  • Natural Language to Workflow: The presenter demonstrated by saying, "Analyze customer reviews from a CSV file. Identify negative sentiment and send alerts to Slack for reviews below three stars."
  • Workflow Components: The generated workflow included:
    • A node for uploading a CSV file with configurable options.
    • A configuration to identify reviews less than three stars.
    • A sentiment analyzer.
    • A block to send alerts to Slack.
  • Code Generation: The "code view" showed that the Gemini model generated code based on these natural language instructions.
  • Execution Example: The presenter provided a negative review, and the system correctly identified it and indicated it would send a Slack message.
• Speech-to-Text System: The presenter mentioned using their own on-device, private speech-to-text system, with a link provided in the video description.

3. Clone of X (Twitter):

• Prompt: Instructions to create a clone of X (formerly Twitter), with a focus on implementing only the front-end. The presenter emphasizes using detailed prompts to ensure accurate instruction following.
• Output: The UI was very close to the current version of X.
  • Specifics: The demo showed posts, including one attributed to Elon Musk and another for OpenAI. Users could like posts.
  • Limitations: Clicking "post" did not perform any action, as only the front-end was implemented. The model did not appear to have access to actual images.

4. Fun Interactive Applications:

• Earth Dashboard:
  • Functionality: Allowed refreshing weather data (simulated, not real-time).
  • Visualization: Featured good visualization capabilities.
  • Limitation: The day/night mode toggle was not fully functional.
• Pixel Pad:
  • Functionality: A drawing application with color selection, eraser, fill, grid toggle, and clear all options.
  • Performance: Described as a "fun little app that actually works."
• Asteroid 2D Game:
  • Functionality: A playable 2D Asteroids-style game.
  • Performance: The presenter confirmed it "works fine."

Testing for Hallucination and Research Capabilities

The presenter tested the model's ability to avoid hallucination and perform research.

1. GPT-6 Announcement Scenario:

• Prompt: The presenter, acting as a tech journalist, provided a fabricated scenario about OpenAI announcing GPT-6 on a specific date and asked the model to perform a web search for more information, including rumored features like memory persistence, on-device reasoning, and neuro-symbolic architecture.
• Execution: The model performed web searches and provided a response.
• Strengths:
  • Used high-credibility links, including the OpenAI website.
  • Correctly identified that GPT-6 had not been announced and would not be released in 2025.
• Weaknesses (Hallucinations/Confusions):
  • Confused the announcement date, stating that "Chat GPT Pulse" was announced on October 29, 2025, when it was announced weeks prior.
  • Mentioned Microsoft increasing its stake to 27% and corporate restructuring, which were true events but not directly related to a GPT-6 announcement.
  • Stated that it was unknown whether the rumored features would be in GPT-6, which is a reasonable statement given the lack of an announcement.
• Conclusion: While the model avoided a major hallucination regarding the GPT-6 announcement, it did confuse some information and timelines. This highlights the importance of fact-checking claims from any agentic system.

2. Mars Colony Executive Report:

• Prompt: "You are an AI research assistant. Your task is to compile a 10-page executive report on the primary technical hurdles for establishing a self-sustaining human colony on Mars."
• Execution: The model generated a plan, sought permission, performed searches, and produced a report.
• Process:
  • The model displayed thinking traces and a plan with tasks that were ticked off as completed.
  • It used reliable and high-quality references.
• Output: A high-quality technical feasibility report.
  • Structure: Followed the requested structure and addressed key challenges like life support, power generation, resource utilization, and habitat construction.
  • Key Findings: Provided detailed findings from its research.
  • Visualizations: Included a workflow diagram and other detailed diagrams, which the presenter speculated might be copied from external sources or generated by the model's image generation capabilities. This ability to incorporate or generate visuals was considered impressive.

3. Interactive 3D Map of Los Angeles:

• Prompt: Create a high-performance, interactive 3D map of Los Angeles as a self-contained HTML file. The map should include specific locations provided by the user, with a "fly to" action triggered by clicking on a location from a sidebar.
• Challenges:
  • Initially attempted to use Mapbox, which is not freely available.
  • When asked to use freely available maps, the implementation did not fully work.
• Partial Success:
  • The "fly to" effect was present when clicking on elements.
  • The presenter suggested a missing layer in the map itself might be the issue, indicating a visualization problem rather than a complete implementation failure.

Minimax M2 Model Details and Performance

The video delves into the technical specifications and performance claims of Minimax M2.

Claims and Benchmarks:

• "Model born for agents and code": Minimax claims M2 is designed for these tasks.
• Cost and Speed: Claimed to be 8% of the price of Claude Sonnet and twice its speed, with free availability for a limited time.
• Coding and Agentic Benchmarks: Positioned as a state-of-the-art open-weight model, approaching the performance of Claude 4.5, and considered the best open-weight model available.
• External Validation: The Artificial Intelligence Index ranked it as the fifth overall best model, not just among open-weight models, which is a significant achievement for its size.

Model Size and Efficiency:

• Parameter Count: 230 billion parameters, with only 10 billion active parameters.
• Sparse Mixture of Experts (SMoE): This architecture contributes to its efficiency, as only a fraction of the parameters are used for inference.
• Comparison:
  • One-third the size of DeepMind's Gopher.
  • Approximately 25% the size of a trillion-parameter model like Mixtral K2.

Speed and Cost:

• Inference Speed: Up to 100 tokens per second on the official API, comparable to Grok 4f (a distilled version of a larger model).
• Efficiency: More efficient in token generation due to its smaller size.
• Cost: Priced comparably to Gemini Flash 2.5, offering performance similar to Sonnet 3.5 or Sonnet 4.
• Price Comparison: Stated as 8% of Claude 3.5 Sonnet's price with nearly double the inference speed. The blog post also compares it to Claude 4.5, though the presenter clarifies the comparison is more accurately with Claude 3.5 Sonnet.

Usage Recommendations:

• Hybrid Approach: Recommended to use Minimax M2 in conjunction with more powerful models like Claude 4.5.
  • Use Claude 4.5 for planning complex tasks.
  • Use Minimax M2 for the actual implementation of well-defined tasks.
• Over-Generation: Open-weight models can sometimes be more verbose and generate more tokens than necessary, which should be considered during comparisons.

Emerging Trends:

• Proprietary Coding Models: The emergence of specialized proprietary coding models like Cognition's Devin 1.5 and Cursor's Composer is noted. There's speculation these might be trained on Chinese-based models, emphasizing the importance of strong open-weight agentic coding models like Minimax M2.

Call to Action:

• The presenter encourages viewers to test the model themselves, rather than relying solely on opinions or benchmarks.
• Minimax M2 and the M2 Agent are available for free trial until November 7th.

Conclusion

Minimax M2, and the agentic system built around it, represents a significant advancement in open-weight language models. Its strengths lie in its impressive instruction-following capabilities, robust agentic performance, and efficient coding abilities. While it has limitations and can still hallucinate, its performance, speed, and cost-effectiveness make it a compelling option for developers and researchers. The ability to generate complex applications, conduct research, and create interactive tools from natural language instructions, coupled with its open nature, positions it as a valuable tool in the AI landscape. The presenter's recommendation to use it in a hybrid approach with more powerful models offers a practical strategy for leveraging its strengths.