Between Kimi K3 and DeepSeek V4, There Is a Time Gap in Native Multimodal Capabilities
Kimi K3 (2.8 trillion total parameters, MoE) and Qwen3.8-Max (2.4 trillion) pursue a "native multimodal" route with early fusion training, while DeepSeek V4-Flash (284 billion) achieves strong Coding/Agent performance without visual capabilities by focusing on post-training optimization Native multimodal means visual and text data jointly shape the base model from pretraining through post-training, enabling "vision in the loop" where agents can inspect screenshots and iteratively refine code out
72
Hot
68
Quality
70
Impact
Analysis
TL;DR
- Kimi K3 (2.8 trillion total parameters, MoE) and Qwen3.8-Max (2.4 trillion) pursue a "native multimodal" route with early fusion training, while DeepSeek V4-Flash (284 billion) achieves strong Coding/Agent performance without visual capabilities by focusing on post-training optimization
- Native multimodal means visual and text data jointly shape the base model from pretraining through post-training, enabling "vision in the loop" where agents can inspect screenshots and iteratively refine code output
- Adding multimodal capability to LLMs creates resource trade-offs: visual data competes with text/code/math for model capacity, potentially degrading existing capabilities unless carefully balanced through early fusion strategies
- The Chinese AI frontier is split on timing and cost: some companies bet on large-parameter native multimodal as essential for long-chain Agent tasks, while others prioritize Coding performance first and defer visual integration
- DeepSeek's V4-Flash scored 1577 on Arena WebDev (near GLM-5.2), proving that post-training alone can significantly boost Coding/Agent abilities without expanding architecture or adding vision, challenging the necessity of immediate multimodal investment
Why It Matters
This article captures a critical strategic inflection point in the Chinese AI industry: whether to invest heavily in native multimodal capabilities now or continue optimizing Coding/Agent performance through post-training alone. For AI practitioners, it highlights the real engineering trade-offs between model capacity, training complexity, and capability balance when adding visual understanding to large language models.
Technical Details
- Kimi K3 architecture: 2.8 trillion total parameters MoE model with ~104 billion active parameters per token, 1M token context window, early fusion pretraining over ~15 trillion mixed image-text tokens, and a newly trained ~400M-parameter visual encoder (MoonViT-V2) that feeds visual representations directly into next-token prediction
- Vision-in-the-loop mechanism: K3 demonstrated the ability to compare target and rendered webpage screenshots, identifying all 5 visual deviations planted by Puter platform without false positives, enabling iterative code refinement based on visual feedback rather than text-only error reports
- DeepSeek V4-Flash approach: 284 billion total parameters (130 billion active per token), no architectural changes or visual capabilities added; performance gains came entirely from post-training improvements, achieving 1577 on Arena WebDev
- Qwen3.8-Max: 2.4 trillion total parameters (950 billion active per token), following the same "large and comprehensive" native multimodal route as K3, supporting visual understanding, Coding, and Cowork capabilities
- Training trade-offs: Kimi K2.5 ablation studies showed that late-stage visual data integration causes temporary text capability degradation before recovery; Apple's MM1 report confirmed that higher resolution and more visual tokens increase both training and inference costs proportionally
Industry Insight
- The "Coding first, multimodal later" strategy (exemplified by DeepSeek) remains viable for near-term commercial competitiveness, but native multimodal may become a differentiator as Agent tasks grow longer and more visually dependent—companies should evaluate their target use cases before committing to resource-intensive multimodal training
- Hiring markets are already shifting: headhunters targeting Moonshot AI (Kimi's developer) specifically seek multimodal expertise post-K3 launch, signaling that multimodal talent is becoming a scarce and strategically valuable resource in the Chinese AI ecosystem
- The fundamental tension between Ilya Sutskever's view (text as sufficient world representation) and Yann LeCun's view (most knowledge exists outside text) remains unresolved in practice; the industry is currently split along these philosophical lines, with timing and cost considerations—not technical consensus—driving current roadmap decisions
Disclaimer: The above content is generated by AI and is for reference only.
Agent Multimodal LLM Code Generation Product Launch
Related Articles
Y Combinator Open-Sources QM: An MIT-Licensed Multiplayer Agent Harness That Runs In Slack And The Web
[GitHub] apache/texera
Anthropic Surpasses OpenAI: The 'Code is King' Logic Behind $965 Billion Valuation
[GitHub] camel-ai/owl
Gemini Agent-to-Agent Attack Method Exposed Secrets, Enabled Pull Request Tampering