Moonshot releases Kimi K3.1 Pro, a 2.8T parameter open-weight model

Moonshot has launched Kimi K3.1 Pro, a multimodal mixture-of-experts model featuring 2.8 trillion total parameters and 104 billion active parameters. The company announced that the model’s weights will be released as open weights on November 4, 2026, marking a significant step toward accessible frontier-level AI infrastructure.

What happened

The new model represents a substantial upgrade in scale and capability over its predecessor, Kimi K3. Moonshot describes K3.1 Pro as a post-trained version of K3, utilizing significantly larger compute resources for reinforcement learning across hundreds of downstream domains. The launch includes both standard and Highspeed API versions, with the latter offering two to three times faster output generation at twice the cost. Additionally, Moonshot introduced Kimi Hosted Agents, a managed environment designed for long-running tasks and multi-agent collaboration.

Independent evaluations place K3.1 Pro firmly among the top tier of current models. Artificial Analysis ranked it fourth on its Intelligence Index with a score of 55, trailing only Opus 5.5, Sonnet 5.5, and Muse Wonder Full. In coding benchmarks, it reached eighth place in the Code Arena WebDev category, improving significantly from its previous ranking. Notably, despite Moonshot’s deliberate removal of cyber-related training data to enhance safety, the model achieved the highest publicly accessible score on the Cyber Index, second only to trusted-access proprietary models.

The release has sparked debate regarding inference economics and benchmark transparency. While token prices have decreased by one-third compared to K3, some independent tests show that higher token consumption per task can offset these savings. Community reactions have been mixed, with enthusiasm for the open-weight promise tempered by concerns over deployment costs and allegations of undisclosed promotional activities. Clarifications have emerged regarding the timeline, confirming the weight release for November 4 rather than earlier rumors.

How it works

Kimi K3.1 Pro utilizes a mixture-of-experts architecture, which allows it to activate only a subset of its parameters for any given input. Specifically, while the model contains 2.8 trillion total parameters, only 104 billion are active during inference. This design aims to balance high capacity with computational efficiency, though it requires substantial infrastructure to manage the large expert pool and runtime state. Deploying such a model locally is not feasible for consumer devices, as the full parameter set would require terabytes of memory even when quantized.

The performance gains stem primarily from extensive post-training reinforcement learning. Moonshot optimized its entire stack, including the RL framework, long-context inference, and sandbox infrastructure, to ensure stable returns from increased compute. A key architectural decision involved removing cyber-related data from mid-training and reinforcement learning phases. This approach aims to reduce latent offensive capabilities at the source rather than relying solely on inference-time refusals, resulting in a reported drop in internal cyber scores while maintaining strong defensive benchmark performance.

Key details

• Model specs: 2.8 trillion total parameters, 104 billion active parameters, and a 1 million-token context window.
• Open weights: Scheduled for release on November 4, 2026, though specific license terms and redistribution conditions remain unspecified.
• Performance: Ranked fourth on Artificial Analysis’ Intelligence Index with a score of 55; eighth in Code Arena WebDev.
• Pricing: Standard API costs $2 per million input tokens and $10 per million output tokens, a one-third reduction from K3.
• Safety: Cyber-related training data was removed, yet the model scored 64% on the public Cyber Index, the highest for an accessible model.
• Speed: The Highspeed API variant offers 2–3× faster output generation but costs twice as much as the standard tier.

Why it matters

For software engineers and technical founders, the arrival of a near-frontier open-weight model changes the calculus for building AI-native products. The promise of open weights allows for greater control over deployment, fine-tuning, and data privacy, which is critical for enterprise applications. However, the distinction between self-hostable and consumer-local is vital. The sheer size of the model means that running it requires significant accelerator infrastructure, making it suitable for cloud-based or dedicated server environments rather than edge devices.

The pricing structure also demands careful analysis. While per-token costs have dropped, the model’s tendency to generate more tokens per task can lead to similar or higher total costs for complex workflows. Developers must evaluate cost per successful task rather than just per-token rates. Furthermore, the model’s strong performance in agentic tasks and coding suggests it could accelerate development cycles, particularly for projects involving long-context understanding or multi-step reasoning.

What you can do

• Monitor the November 4, 2026 release date for the open weights and review the license terms for commercial use restrictions.
• Benchmark your current workloads against K3.1 Pro using the Highspeed API to assess if the speed premium justifies the cost for your specific latency requirements.
• Evaluate the model’s long-context capabilities for projects requiring full-project code understanding or extensive document analysis.
• Consider the implications of the cyber-safety approach for your security posture, noting that high defensive scores do not guarantee unrestricted offensive capability.
• Test the model in your specific domain, such as 3D modeling or financial analysis, to verify vendor claims against your own quality standards.
• Plan infrastructure upgrades if you intend to self-host, accounting for the massive memory requirements of the 2.8 trillion parameter model.