Kimi K3.1 Pro has beaten GPT-6 Astra and Claude Fable 5.1 in a key AI ranking. Moonshot's new model got 55 points on the Artificial Analysis Intelligence Index, while its two rivals each got 53. Quite a result, given how OpenAI and Anthropic have tried to curb model distillation in China.
The top spots still belong to Claude Opus 5.5, Claude Sonnet 5.5, and Meta's Muse Wonder Full. Meanwhile, Code Arena has Kimi in eighth place, level with Claude Opus 5. Not bad for a model that many thought would struggle.
Moonshot kept parts of its older Kimi K3 design but put far more compute into training after the base phase. The focus is on 3D work, complex finance tasks, and AI agents. Those gains come with some serious specs.
For scale, the model has 2.8 trillion parameters, yet it uses just 104 billion per token. Its stated activation rate is 1.8%, though that figure does not match the parameter counts. A native one-million-token context limit is also part of the package.
One neat detail is NoPE, which drops the usual position data used by RoPE. Moonshot uses this method across its network. The aim is to handle long text with less compute and better general use.
On the hardware side, MXFP4 handles weights, while MXFP8 deals with activations. Special training helps the model work with these low-bit formats. API fees sit at $2 for input, $10 for output, and $0.20 for cache hits, each per million tokens. All three rates are down 33% from Kimi K3.
And the model's weights are due for public release. That gives the AI crowd another chance to test Moonshot's claims outside the firm's lab.
The main tech change worth a look is Kimi Delta Attention, or KDA. Its setup blends two ways to handle text. Three stages use recurrent memory, while a fourth uses Gated Multi-Head Latent Attention to work with a wider set of past tokens.
That mix helps keep the KV cache small. In a standard system, past token data takes up more space as prompts grow. Recurrent memory avoids much of that growth, but fine details can still get lost. The other attention stage helps reduce that risk.
Moonshot's stated layout has 69 layers, with 24 gated global attention layers in the mix. The goal is clear. Long prompts need less memory, while the model still gets access to broader context when a task calls for it.
There is a safety issue here as well. Moonshot says it cut cyber-related training data from both its middle training phase and reinforcement learning. Before those cuts, an unredacted model had passed 90.5% of internal exploit tests tied to the Linux kernel, Node.js, and Exim.
Despite that change, Kimi K3.1 Pro still got 81% on xdev-bench-hard. The result suggests that its broader coding skills stayed strong, at least on that test. Independent checks will give a better view of how well those scores hold up.
The top spots still belong to Claude Opus 5.5, Claude Sonnet 5.5, and Meta's Muse Wonder Full. Meanwhile, Code Arena has Kimi in eighth place, level with Claude Opus 5. Not bad for a model that many thought would struggle.
Moonshot kept parts of its older Kimi K3 design but put far more compute into training after the base phase. The focus is on 3D work, complex finance tasks, and AI agents. Those gains come with some serious specs.
For scale, the model has 2.8 trillion parameters, yet it uses just 104 billion per token. Its stated activation rate is 1.8%, though that figure does not match the parameter counts. A native one-million-token context limit is also part of the package.
One neat detail is NoPE, which drops the usual position data used by RoPE. Moonshot uses this method across its network. The aim is to handle long text with less compute and better general use.
On the hardware side, MXFP4 handles weights, while MXFP8 deals with activations. Special training helps the model work with these low-bit formats. API fees sit at $2 for input, $10 for output, and $0.20 for cache hits, each per million tokens. All three rates are down 33% from Kimi K3.
And the model's weights are due for public release. That gives the AI crowd another chance to test Moonshot's claims outside the firm's lab.
The main tech change worth a look is Kimi Delta Attention, or KDA. Its setup blends two ways to handle text. Three stages use recurrent memory, while a fourth uses Gated Multi-Head Latent Attention to work with a wider set of past tokens.
That mix helps keep the KV cache small. In a standard system, past token data takes up more space as prompts grow. Recurrent memory avoids much of that growth, but fine details can still get lost. The other attention stage helps reduce that risk.
Moonshot's stated layout has 69 layers, with 24 gated global attention layers in the mix. The goal is clear. Long prompts need less memory, while the model still gets access to broader context when a task calls for it.
There is a safety issue here as well. Moonshot says it cut cyber-related training data from both its middle training phase and reinforcement learning. Before those cuts, an unredacted model had passed 90.5% of internal exploit tests tied to the Linux kernel, Node.js, and Exim.
Despite that change, Kimi K3.1 Pro still got 81% on xdev-bench-hard. The result suggests that its broader coding skills stayed strong, at least on that test. Independent checks will give a better view of how well those scores hold up.
