Kimi K2 — More than a Coder

Key Concepts

• Kimik2: An excellent coding model with a state-of-the-art deep researcher.
• Deep Researcher: A single autonomous agent for multi-turn search and reasoning.
• Agentic Systems: Systems employing autonomous agents for specific tasks.
• Reinforcement Learning (RL): Training method used to improve the agent's research capabilities.
• Context Management: The agent's ability to retain important information and discard irrelevant data.
• Quantization: A technique to reduce the precision of numerical values in a model, potentially affecting inference accuracy.
• API Providers: Platforms hosting Kimik2 and other models for inference.
• Tokens per Second: A measure of the speed at which a model generates text.

1. Kimik2 and Deep Researcher Overview

• Kimik2 has a top-tier deep researcher, released almost a month ago, which was tested against other deep researchers.
• The deep researcher is a single autonomous agent excelling in multi-turn search and reasoning.
• It uses three tools in parallel: real-time internal search, a text-based browser for web tasks, and a coding tool for automated code execution.
• The system emphasizes context management, a crucial aspect of agentic systems.

2. Agentic Systems Debate

• There's an ongoing debate about single-agent vs. multi-agent systems.
• Cognition (creators of Devin) advocates for single-agent systems, especially for coding tasks.
• Anthropic recommends multi-agent systems for search-related tasks.
• Kimik2's researcher uses a single-agent system for search, enabling long-horizon tasks (up to 23 reasoning steps and 200 URLs per task).

3. Performance and Training

• The Kimik2 researcher was state-of-the-art a month ago, outperforming other deep researchers in humanity's last exam, except for the new Grok-4.
• It was trained end-to-end using reinforcement learning specifically for research purposes.
• This approach differs from multi-agent workflows, training a single model to solve problems holistically.
• The agent explores strategies, receives rewards for correct solutions, and learns from full trajectories.
• Performance improves as the agent takes more steps during training.

4. Emerging Agentic Capabilities

• Reinforcement learning led to new agentic capabilities:
  • Resolving conflicting information through iterative hypothesis refinement and self-correction.
  • Demonstrating caution and rigor, performing additional searches and cross-validating information before answering.
• The agent seems to recover from context rot easily and is cautious due to its training for search.

5. Context Management Details

• Long research trajectories can lead to massive context, exceeding limitations within 10 iterations.
• Kimik2's agent uses context pruning, retaining important information and discarding unnecessary documents.
• This allows up to 50 iterations and 30% more reliable iterations, enabling longer horizon tasks.

6. Experiment: API Provider Comparison

• The experiment aimed to determine if different API providers host Kimik2 at different quantization levels and if this impacts inference accuracy.
• The task involved data collection on API providers, quantization levels, tokens per second, context window, pricing, and code examples.
• The task was given to Gemini, OpenAI 3 (with deep research), Grok (with deep search), Kimik2, Perplexity, Claude (with search), and Menus.

7. Results from Different Models

• Gemini: Listed official Moonshot, Grok, Deep Infra, Fireworks, Together AI, Open Router, and Hugging Face. Provided correct information and a hypothetical code generation table.
• OpenAI 3: Similar providers listed, including Silicon Flow/Cloud. Provided quantization levels and tokens per second (potentially hallucinated).
• Grok Deep Search: Listed information without a table, found Nova as an additional provider, but didn't find quantization information.
• Perplexity: Consistent pricing information, but no quantization information.
• Claude Sonnet: Listed information without a table, included Runpod (not an API provider).
• Menus: Listed Grok, Together AI, Fireworks, Moonshot, Parasail, missing Deep Infra.
• Kimik2: Used Chinese web links, provided detailed information on providers, URLs, pricing, and capabilities. Created an interactive website report with an executive summary, model details (tokens used, context window, optimizer), and provider-specific information.

8. Kimik2's Unique Features

• Kimik2 uses Chinese web links in its research, providing access to more information.
• It generates an interactive website report, well-formatted compared to other deep researchers.
• The report includes an executive summary, model details, and provider-specific information.

9. Observations and Future Work

• Different providers may host Kimik2 at different quantization levels.
• The tokens per second information from OpenAI might be hallucinated.
• A subsequent video will test different providers to assess the impact of quantization on inference accuracy.

10. Conclusion

• Kimik2's deep researcher is a powerful tool with unique features like Chinese web link integration and interactive reports.
• Reinforcement learning has enabled emerging capabilities like resolving conflicting information and demonstrating caution.
• Context management allows for longer research trajectories.
• Gemini is recommended for free deep searches, and Kimik2 is noted for being less agreeable, which is a positive trait.