Kimi K3 vs. GPT and Claude: Can This Open-Weight Giant Really Win?
All blog articles
Two point eight trillion parameters. That's the number Moonshot AI just put on the table with Kimi K3, and it instantly makes every other open model on the planet look modest.
China's Kimi K3 Sets a New Open-Source Size Record
Moonshot AI, the Beijing startup behind the Kimi chatbot, released Kimi K3 on July 16 with a headline number nobody in open-source AI had touched before. Reports had pointed to a parameter count between 2 and 3 trillion, and Moonshot landed at 2.8 trillion, comfortably ahead of DeepSeek's V4-Pro at 1.6 trillion.
Xiaomi and Z.AI, the other two Chinese labs racing for open-weight bragging rights, sit at 1.02 trillion and 744 billion respectively. Kimi K3 simply lapped them.
A Million Tokens and a New Attention Trick
Size alone would be a gimmick without something to back it up, so Moonshot paired the parameter jump with a 1,048,576-token context window. That means feeding the model an entire codebase, a legal contract stack, or a full novel trilogy in one shot.
The engineering behind it is called Kimi Delta Attention, a hybrid linear attention mechanism combined with Attention Residuals, built to keep that huge context usable without melting your GPU budget. Native visual understanding is baked in too, so K3 reads screenshots, logs, and video the same way it reads text.
Where Kimi K3 Lands Next to Claude and GPT
Artificial Analysis Intelligence Index v4.1
Overall score — higher is better
| Claude Fable 5 (with fallback) | 60 | |
| GPT-5.6 Sol (max) | 59 | |
| Kimi K3 | 57 | |
| Claude Opus 4.8 (max) | 56 | |
| GPT-5.6 Terra (max) | 55 | |
| GPT-5.5 (xhigh) | 55 | |
| Grok 4.5 (high) | 54 | |
| Claude Sonnet 5 (max) | 53 | |
| GPT-5.6 Luna (max) | 51 | |
| KLM-5.2 (max) | 51 | |
| Muse Spark 1.1 (xhigh) | 51 | |
| Gemini 3.5 Flash | 50 |
Source: Artificial Analysis
Moonshot's own claim is bold: K3 ranks second in overall intelligence, only behind Claude Fable 5 and GPT-5.6 Sol, and ahead of Gemini on the company's internal evaluation set. Independent benchmarks are now available. Artificial Analysis gives Kimi K3 a score of 57 on its Intelligence Index v4.1, ranking it among the top-performing models on the market. This result confirms that K3 rivals the best proprietary models, even if it does not necessarily validate all the rankings put forward by Moonshot AI.
Pricing tells its own story: $0.30 per million cached input tokens, $3 for a cache miss, and $15 for output. That's pocket change next to what OpenAI and Anthropic charge for their flagship models, echoing the pricing gap already visible when OpenAI's GPT-5.6 Sol shipped under restricted access.
Why an Alibaba-Backed Startup Keeps Winning This Race
Moonshot isn't some garage operation. Backed by Alibaba and Tencent, the company was valued near $20 billion in May and is reportedly raising fresh capital that could push it toward $31.5 billion.
This is the ninth time in roughly a year that a Kimi release has reset the open-source scale record, following a lineage that started at K2's 1 trillion parameters. If bigger keeps winning, DeepSeek, Zhipu, and every Western lab watching the open-weight scoreboard just got a new number to beat.
How many parameters does Kimi K3 have?
Kimi K3 has 2.8 trillion total parameters, which Moonshot AI says makes it the largest open-source model released to date, ahead of DeepSeek's V4-Pro at 1.6 trillion.
Is Kimi K3 free to use?
Kimi K3 is not free but is priced far below top closed models, at $0.30 per million cached input tokens, $3 for uncached input, and $15 per million output tokens.