April 29, 2026

Kimi is the AI assistant and model family built by Moonshot AI (月之暗面, Yuè zhī àn miàn — “Dark side of the moon”), a Chinese AI startup that built its early reputation almost entirely on one technical differentiator: context length.

The Beginning (2023)

Moonshot AI was founded in March 2023 by Yang Zhilin, Zhang Yuhui, and a small team mostly from top Chinese universities and research labs. Yang Zhilin, the CEO, had previously been a researcher at Tsinghua and Google Brain, and was the co-author of Transformer-XL and XLNet — two influential pre-BERT architectures that shaped early large model research.

The company raised a large seed round quickly and attracted attention from investors including HongShan (formerly Sequoia China) before having a public product.

In October 2023, Moonshot launched Kimi Chat, a conversational assistant notable for supporting a 200K token context window — among the largest offered publicly at the time. While other labs were racing on benchmark scores, Moonshot bet on context as the primary axis of differentiation.

Long Context as a Strategy (2024)

2024 was the year Kimi’s bet on context length paid off in terms of visibility.

• March 2024: Kimi updated to support up to 2 million tokens — a claim that drew widespread attention globally. The product framing was simple: feed it an entire codebase, a long document corpus, or hours of transcripts. Whether users needed that much context in practice was secondary to establishing the capability.
• Spring 2024: Moonshot AI closed a major funding round with participation from Alibaba, reaching a multi-billion dollar valuation. The investment signaled that Alibaba was hedging its bets beyond its own Qwen models.
• Mid-2024: Kimi gained multimodal capabilities, adding image understanding alongside text. Kimi Explorer (探索版) was introduced as a research-oriented mode with real-time web search and multi-step reasoning, broadly similar to Perplexity’s approach but integrated directly into the Kimi product.
• By the second half of 2024, Kimi had become one of China’s most-used AI assistants, competing directly with Baidu’s ERNIE Bot and ByteDance’s Doubao.

Reasoning and Open Weights (2025)

• January 2025 — Kimi k1.5: Moonshot released its first dedicated reasoning model, accompanied by a technical report. The notable claim was methodological: rather than relying on Monte Carlo Tree Search or process reward models (the approach associated with DeepSeek-R1 and similar models), Kimi k1.5 used long-context scaling — training the model to reason by extending its context over a long chain of thought rather than through structured search. The model showed competitive results with OpenAI’s o1 on math and coding benchmarks, and handled both text and images.
• June 2025 — Kimi k2: An open-weight model released on Hugging Face, positioned explicitly for agentic use cases — tool use, multi-step planning, coding. It uses a mixture-of-experts architecture with roughly 1 trillion total parameters and around 32 billion active per forward pass. At release, it ranked among the top open-weight models on coding benchmarks and drew comparisons to DeepSeek-V3 in capability and efficiency.

What distinguishes Kimi’s trajectory from peers like Qwen or Baidu is the consistent focus on a single architectural claim — long context — as the product’s identity, even as the rest of the field caught up on that dimension. The shift toward reasoning (k1.5) and then agentic models (k2) follows a similar pattern: pick one clear differentiator, build around it, and iterate.