Chinese AI Startup Expands Model Context but Still Trails DeepMind’s Gemini in 2025 Race

China's latest AI contender just upped its context game—but still plays catch-up to DeepMind's heavyweight champ.

The Context Gap

While pushing boundaries with expanded model context, the startup hits a familiar wall: Gemini's established dominance. They're scaling up, but not closing in.

Active Development, Passive Results

Teams work round-the-clock, deploying cutting-edge architectures and aggressive training regimens. Yet the performance delta remains stubbornly wide—like bringing a knife to a quantum fight.

Finance Bite: Probably burned through another round of VC funding just to prove they're still in the game. Because nothing says 'innovation' like chasing your competitor's shadow.

Bottom line: Expansion signals ambition, but the leaderboard doesn't lie—and neither do the benchmarks.

TLDRs;

• Moonshot AI doubles Kimi K2 context window to 256K tokens but lags behind DeepMind’s 1M context Gemini model.
• The update improves coding and writing skills, but lacks reasoning and vision features, limiting its competitive edge.
• API issues delayed beta testing, with access restricted to just 20 Discord users, reflecting scaling challenges.
• Despite a $3.3B valuation and Alibaba-Tencent backing, Moonshot still trails global leaders in key AI capabilities.

Moonshot AI, a Beijing-based artificial intelligence startup backed by Alibaba and Tencent, has revealed an expanded version of its flagship Kimi K2 model.

The new update increases the model’s context window from 128,000 to 256,000 tokens, improving its capacity to process larger volumes of text and code. According to growth team member Aspen Choong, the upgrade also brings stronger coding and creative writing capabilities, while reducing the risk of hallucinations.

Despite these improvements, the company has decided against including advanced reasoning or vision features in this release. Technical hiccups have already delayed beta testing, with the rollout on Discord limited to a small group of just 20 testers due to “API issues.”

Still Behind Global Leaders

While doubling the context size marks progress for the young company, the update underscores the gap between Chinese AI developers and global leaders.

Moonshot AI’s founder, Yang Zhilin, admitted that “256K context is simply not enough as you need millions or even more” to remain competitive.

For context, Google DeepMind’s Gemini 2.5 Pro already offers a staggering 1 million-token context window, four times the size of Moonshot’s new release. The difference is significant. Larger context windows enable AI models to analyze long documents, maintain coherence across extended conversations, and manage complex codebases more effectively.

This technological gap highlights a persistent challenge for Chinese AI startups: while they are making rapid strides, their progress still lags behind leading Western labs in areas fundamental to long-term competitiveness.

High Valuation, Higher Expectations

Founded in March 2023, Moonshot AI has quickly climbed to a US$3.3 billion valuation, supported by heavyweight investors Alibaba and Tencent.

Its Kimi K2 model has also gained industry recognition, ranking eighth overall and fourth in coding performance on the developer leaderboard LMArena.

Kimi K2-0905 update 🚀
– Enhanced coding capabilities, esp. front-end & tool-calling
– Context length extended to 256k tokens
– Improved integration with various agent scaffolds (e.g., Claude Code, Roo Code, etc)

🔗 Weights & code: https://t.co/83sQekosr9
💬 Chat with new Kimi… pic.twitter.com/mkOuBMwzpw

— Kimi.ai (@Kimi_Moonshot) September 5, 2025

However, with such strong financial backing and high visibility, expectations for Moonshot are considerable. Analysts argue that modest context window gains and testing delays risk reinforcing the perception that Chinese AI firms are catching up rather than setting the pace.

Technical Roadblocks and Market Pressures

The delayed beta rollout due to API complications is not unique to Moonshot AI, but it raises questions about scalability and reliability. In the AI race, infrastructure performance is just as critical as algorithmic breakthroughs. As companies increasingly deploy models in real-world applications, technical reliability becomes a deciding factor in maintaining competitive momentum.

Moonshot’s choice to restrict initial testing to a small group of users reflects a cautious strategy, aimed at minimizing reputational damage should further issues arise. Yet, in a market where global giants are racing ahead with more powerful models, delays risk slowing momentum and frustrating early adopters eager to test new features.

Conclusion

Moonshot AI’s latest Kimi K2 upgrade signals progress in China’s rapidly evolving AI industry, but the advances remain incremental compared to breakthroughs achieved by competitors like DeepMind.

With $3.3 billion in backing and a strong ranking in developer benchmarks, Moonshot has proven it can play in the global field. But unless it closes the widening gap in Core capabilities such as context scaling, the company risks being remembered as a fast follower rather than a market leader.