Moonshot Unveils Kimi K3.1 Pro With 2.8 Trillion Parameters and Agentic Gains
Moonshot has announced Kimi K3.1 Pro, a massive 2.8T-parameter mixture-of-experts model featuring a 1-million-token context window and strong agentic improvements.
Moonshot has officially launched Kimi K3.1 Pro, a multimodal, 3T-class mixture-of-experts model featuring 2.8 trillion total parameters and 104 billion active parameters. Alongside the model release, Moonshot introduced a 1-million-token context window, standard and high-speed API options, and Kimi Hosted Agents for long-running workflows. Independent evaluations show significant capability jumps, though the release has sparked discussions regarding serving costs, safety interventions, and deployment realities.
Performance and Independent Benchmarks
Independent testing from Artificial Analysis placed K3.1 Pro fourth on its Intelligence Index with a score of 55, trailing Opus 5.5 at 58, and Sonnet 5.5 and Muse Wonder Full at both scoring 56. The model's improvements heavily favor agentic tasks rather than uniform knowledge gains. In coding evaluations, the model climbed to #8 in Code Arena WebDev with 1,692 points, marking a nine-place improvement over its predecessor, Kimi K3, and returning Moonshot to the Arena's top four labs.
On the cybersecurity front, Moonshot intentionally removed cyber-related data during mid-training and reinforcement learning rather than relying solely on inference-time refusals. Despite this intervention, K3.1 Pro scored 64% on Artificial Analysis' Cyber Index, marking the highest publicly accessible result and placing second overall behind trusted-access GPT-6 Sol Daybreak Blue at 69%. Artificial Analysis reported zero safety refusals on its tasks for the model.
Standard pricing for K3.1 Pro is set at $2 per million input tokens, $10 per million output tokens, and $0.20 per million cached-input tokens, which represents a one-third reduction compared to Kimi K3. A Highspeed version is also available, generating outputs two to three times faster at twice the standard pricing. While token prices are lower, task-specific cost comparisons vary depending on token consumption per workload.
What it means for developers
Developers seeking flexible, cost-effective infrastructure can try top AI models cheaply through one API at https://apixoai.online. With Kimi K3.1 Pro entering the market alongside other frontier systems, engineering teams gain access to high-parameter mixtures of experts capable of handling full-project coding, financial modeling, and CAD tasks. However, developers must factor in serving economics, as deploying models with active parameter pools requires robust accelerator infrastructure rather than standard consumer-device-local setups. Furthermore, harness dependencies and varying provider implementations mean testing specific workflows remains essential for optimal performance.
Release Status and Community Reception
While the API launched immediately, open weights were explicitly promised for November 4, 2026. The initial rollout generated considerable community discussion, including debate over benchmark presentation, deployment economics, and comparisons with existing proprietary and open-weight alternatives. Observers noted that while the model narrows the gap with closed-AI systems, practical deployment will depend heavily on real-world inference costs and the exact licensing conditions accompanying the upcoming weight release.
Source: Kimi 3.1 Pro 104BA-2.8T: Fable-5.1-Class Open Weights narrows Closed-AI lead to 1 month... and Pro Highspeed API with 2-3× speed — Latent Space. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
