Moonshot AI targets a $50 billion valuation as it negotiates its final pre-IPO financing round. The company plans to complete its offshore red chip structure adjustment by the end of October 2026 to prepare for a potential Hong Kong listing later this year. Insiders say a current financing round values the Beijing-based lab at approximately $31.5 billion. This capital injection follows intense market interest. Moonshot AI is part of China’s "AI Tigers" alongside DeepSeek, Qwen, and MiniMax. The company functions as a provider of AI-based website optimization solutions for e-commerce businesses and received $10 million in seed funding. While the lab grew quickly after its 2023 founding, its current goal involves completing a final capital injection before listing. The company is expected to complete a round of financing launched this summer within the next few days. Does the company prioritize valuation over immediate profitability?
Model Lifecycle and API Economics
The Kimi model landscape changed when Moonshot AI retired the K2.5 and the entire legacy V1 family on August 31, 2026. This retirement caused 404 errors for unmaintained integrations that previously relied on K2.5’s $0.60 input and $3.00 output rates. Developers now migrate workloads to the K2.6 model or the coding-specific K2.7 Code. The K2.6 model, which uses a 1-trillion-parameter Mixture-of-Experts architecture that activates only 32 billion parameters per request, supports a 256K-token context window that exceeds the 128K capacity of OpenAI’s GPT-4o by a significant margin. K2.6 also includes the Agent Swarm feature that coordinates up to 300 parallel sub-agents to execute 4,000 coordinated steps. In a real test, K2.6 independently iterated through 12 optimization strategies over 13 continuous hours, made over 1,000 tool calls, and modified more than 4,000 lines of code. This version also includes Claw Groups as a research preview, which lets humans and agents share work at runtime. The model achieved a 185% leap in median throughput. On BrowseComp Swarm, K2.6 achieves an 86.3% score compared to the 78.4% score of K2.5.
Kimi K3, released July 16, 2026, is the reasoning-and-agents tier. K3 is a 2.8-trillion-parameter multimodal flagship. K3 costs $3.00 per million input tokens and $15.00 per million output tokens. While the K2 line uses a 256K context window, the K3 flagship provides a 1,048,576-token context window. K3 also includes a 90% discount on cache hits, which costs $0.30 per million input tokens. The K2.7 Code variant maintains the same $0.95 input and $4.00 output rates as K2.6, though it dropped the MoonViT vision encoder. K2.7 Code HighSpeed provides a higher output rate of 180 tokens per second for $1.90 per million input and $8.00 per million output tokens.
| Model | Context Window | Input Price (per 1M) | Output Price (per 1M) |
|---|---|---|---|
| K2.6 | 256K | $0.95 | $4.00 |
| K2.7 Code | 262,144 | $0.95 | $4.00 |
| K2.7 HighSpeed | 262,144 | $1.90 | $8.00 |
| K3 | 1,048,576 | $3.00 | $15.00 |
Technical Capacity and Privacy
Moonshot AI stores API data in Singapore and trains on submitted content by default with no documented opt-out. The company also lacks a published SOC 2 report or a signed HIPAA BAA. For teams requiring data residency, self-hosting K2 models on H200 or B200 clusters provides an alternative. A production 8x H200 SXM5 cluster running FP8 costs about $5.21 per million tokens on-demand. An 8x B200 SXM6 cluster running AWQ INT4 provides roughly 1.8x the throughput of the H200 cluster. The K2 architecture uses 8-way tensor parallelism within a single node. The K2.6 model’s 256K-token context window covers approximately 200,000 words of English text. This capacity is larger than the 128K capacity of GPT-4o and larger than the 200K context window of Claude 3.5 Sonnet. K2.6 also ranks 26th out of 115 models on multimodal benchmarks. For the K2 architecture, the router sends any token to any of the model’s experts, so every expert’s weights sit resident in memory regardless of how many fire. A budget 4x B200 SXM6 node running AWQ INT4 provides 768GB of total VRAM, while an 8x H200 SXM5 node provides 1,128GB of HBM3e memory. If you manage high-volume API traffic, do you find the cost of self-hosting an H200 cluster more efficient than the K2.6 API?
