Moonshot AI Launches Kimi K3.1 Pro with 3T Parameters, Outperforms GPT-6 Astra and Claude Fable 5.1

Moonshot AI Unveils Flagship Kimi K3.1 Pro Model

Moonshot AI introduced Kimi K3.1 Pro, a flagship large language model that has emerged as the strongest model out of China, challenging expectations after OpenAI and Anthropic restricted Chinese AI labs from distilling their leading models2

2

. The large language model is a post-trained upgrade of Kimi K3, released in July, comprising 896 individual neural networks with 3 trillion parameters, though it activates only 16 networks at any given time to minimize inference costs1

1

. The model features 2.8 trillion total parameters with 104 billion active parameters per token, representing approximately 1.8 percent activation ratio, and supports a 1 million token native context window2

2

.

Advanced Architecture with Kimi Delta Attention

Kimi K3.1 Pro preserves architectural elements from its predecessor while utilizing vastly scaled-up post-training compute framework optimized for 3D synthesis, advanced financial modeling, and agentic workflows2

2

. The model features Kimi Delta Attention, a custom implementation of the memory-saving technology Gated DeltaNet that reduces RAM usage in the attention mechanism when interpreting user prompts1

1

. Unlike conventional quadratic-attention approaches where KV cache scales linearly with context, Kimi Delta Attention adopts a hybrid linear-attention mechanism using Recurrent and Hybrid Interleaving states in a 3:1 ratio, dramatically reducing KV cache requirements2

2

. The model utilizes NoPE (No Positional Embeddings) throughout the entire network, completely omitting traditional RoPE layers, which facilitates a more flexible and larger context window, better generalization, and improved computational efficiency2

2

.

Superior Performance Across Multiple Benchmarks

Kimi K3.1 Pro scored 55 on the Artificial Analysis Intelligence Index, surpassing GPT-6 Astra and Claude Fable 5.1, which both scored 53, and ranking just below Claude Opus 5.5, Claude Sonnet 5.5, and Meta's Muse Wonder Full2

2

. The model excels particularly in 3D modeling, scoring 92.7% on the BenchCAD benchmark, placing it within four percentage points of GPT-6 Astra and Claude Opus 5.51

1

. On Terminal-Bench 4.0, a coding benchmark, Kimi K3.1 Pro achieved 65.7%, outperforming GPT-6 Astra by more than 7% though trailing slightly behind Claude Opus 5.51

1

. The model outperformed both GPT-6 Astra and Claude Opus 5.5 on OfficeQA Pro v2, which measures complex financial analysis capabilities1

1

. It also ranks 8th on Code Arena, matching Claude Opus 52

2

.

3D Modeling and Animation Capabilities

Kimi K3.1 Pro can generate static 3D models and animate them by writing code snippets that specify geometry in numerical terms rather than drawing1

1

. Internal testing demonstrated the model generating animated simulations of human motion and creating multiple 3D video games, tasks considered difficult for AI models due to the large amount of code required and the risk of bugs causing crashes1

1

.

Cybersecurity Trade-offs and Safety Measures

Kimi K3.1 Pro's performance on cybersecurity benchmarks significantly trails GPT-6 Astra and Claude Opus 5.5 scores1

1

. Moonshot AI revealed it "identified and removed cyber-related data" from the model's post-training dataset to address safety risks2

2

. The company plans to equip future models with more advanced vulnerability analysis features1

1

.

Enhanced Infrastructure and Efficiency

Moonshot AI's post-training workflow for Kimi K3.1 Pro consumed significantly more processing power than similar past projects and ran on an upgraded post-training stack with automated quality checks and boosted infrastructure efficiency1

1

. The model employs MXFP4 weights with MXFP8 activations achieved through quantization-aware training for massive hardware efficiency2

2

.

Pricing and Open-Sourcing Model Weights

Kimi K3.1 Pro is priced at $2/$10/$0.20 per 1 million input/output/cache hit tokens respectively, representing 33 percent lower rates than Kimi K3 across all three categories2

2

. The model is available via Moonshot AI's paid cloud services at launch, with the company planning to open-source the model's weights by November 41

1

.