Moonshot AI's Kimi K3.1 Pro opens with 2.8 trillion parameters, topping open model rankings.

Today, Moonshot AI open-sourced the Kimi K3.1Pro model weights, ranking first among open models on the Agent Arena leaderboard with a net improvement score of +5.1%, leading in productivity, programming, and chat tasks. The model employs a 2.8 trillion parameter mixture-of-experts architecture, supports a 1 million token context window, and has been optimized across multiple domains through reinforcement learning. Although the original weights are 5.6 TB in size, making local deployment challenging, it achieved a score of 65.7% on terminal agent benchmarks, approaching the performance of top-tier closed-source models.

Article author, source: AIBase

On November 5, in accordance with a commitment made on October 24, Moonshot AI officially released the model weights of Kimi K3.1Pro on the Hugging Face platform. With this official open-sourcing of the weights, the model has performed exceptionally well on the Agent Arena leaderboard operated by Arena, ranking first among all open-weight models, 13th overall among all evaluated models, and securing the top position in all three core categories: productivity, programming, and chat.

Unlike traditional pairwise blind tests, Agent Arena features a unique evaluation method. Instead of relying on users to vote between two responses, it deeply analyzes real user-agent conversations to automatically extract multiple key signals, including explicit user confirmations, praise and complaints, correction instructions, and Shell command execution errors. Based on this, the leaderboard scientifically estimates the net improvement rate of each model relative to the average using a randomized inference method called "causal tracing."

This impressive ranking for Kimi K3.1Pro is built on over 17,500 real-world, long-term agent task conversations from users worldwide. Its overall net improvement rate reaches +5.1%, a significant 4.2 percentage point increase compared to the +0.9% of its predecessor, K3 (Max).

From the specific sub-indicators, all core capabilities have undergone comprehensive evolution:

Bash Reliability: A dramatic increase from -4.5% (ranked 13th) in the previous generation to +7.9%, now ranking first among all open models.

Praise vs Complaint (User Satisfaction Metric): Steadily improved from +1.3% to +7.7%.

Steerability (direction adjustment and correction capability): Increased by +6.5%, ranking highest among open models.

Confirmed Success (Task Completion Confidence): Increased by +3.9%, ranking fourth among open models.

In terms of technical architecture, Kimi K3.1Pro employs a Mixture of Experts (MoE) architecture, with a total of 2.8 trillion parameters, activating approximately 104 billion parameters during a single inference, and supporting an ultra-long context window of about 1 million tokens.

Compared to its predecessor, the K3, the K3.1Pro has undergone extensive training via collaborative reinforcement learning, covering hundreds of downstream application domains. Additionally, it innovatively replaces some full attention layers with gated sparse attention layers, effectively reducing computational overhead for long-context modeling while maintaining strong performance.

Although the model weights have been successfully released to the open-source community, there remains a practical hardware barrier between “downloadable” and “locally runnable.” The weight files, stored at BF16 precision, amount to approximately 5.6 TB for 2.8 trillion parameters; even with INT4 quantization, around 1.4 TB of GPU memory is still required, making local deployment on consumer-grade single GPUs essentially infeasible.

On the authoritative Terminal-Bench4.0 benchmark, K3.1Pro achieved a score of 65.7%. This result not only surpasses GPT-6Astra’s 58.2%, but is also within less than 1 percentage point of Claude Opus 5.5’s 66.4%. With the full open-sourcing of these weights, Moonshot is injecting fresh energy into the global open-source large model ecosystem through continuous technical sharing.