Moonshot launches Kimi K3.1 Pro, a 2.8T-parameter model, with open weights promised Nov. 4

Moonshot AI has launched Kimi K3.1 Pro, a multimodal mixture-of-experts model that it calls "3T-class." It has 2.8 trillion total parameters, 104 billion active parameters and a 1-million-token context window. Moonshot lists full-project coding, Blender and CAD work, financial modeling and long-running agents as its main strengths.

Standard and Highspeed versions are available on Kimi's website and app, Kimi Work, Kimi Code and the API. Moonshot says Highspeed keeps the model's intelligence while generating output two to three times faster. The official-account summary says it costs twice as much. The launch also includes Kimi Hosted Agents, a managed harness with a 24/7 execution environment for long-running work and multi-agent collaboration.

Moonshot describes K3.1 Pro as a post-trained version of Kimi K3 with substantially larger post-training compute. It says it optimized the reinforcement-learning framework, long-context inference, sandbox infrastructure and rollout scheduling to support joint RL across hundreds of domains. Little other research detail was disclosed.

Independent results put the model near the frontier. Artificial Analysis ranks it fourth on its Intelligence Index with a score of 55, behind Opus 5.5 at 58 and Sonnet 5.5 and Muse Wonder Full at 56, and says the gains were strongest on agentic tasks. Arena ranks it #8 in Code Arena: WebDev, and Moonshot returns to Arena's top four labs. On Artificial Analysis' Cyber Index it scored 64%, the highest publicly accessible result, behind only the trusted-access GPT-6 Sol Daybreak Blue at 69%, despite Moonshot reducing cyber-related training.

If the weights arrive as promised, an open model would sit close to the best closed systems. Latent Space frames this as narrowing the closed-model lead to about a month. That is the newsletter's characterization rather than a measured figure.

The model's size limits who can use it. Latent Space's illustrative arithmetic puts 2.8T parameters at about 5.6 TB at 16-bit or 1.4 TB at 4-bit, before overhead. Active parameters do not reduce the storage needed for the full expert pool. Commentator giffmana warned against calling trillion-parameter models "local," since self-hostable and consumer-device-local are different things.

The main open question is the weights release. Latent Space says the coverage does not establish the license, redistribution or commercial-use conditions, checkpoint precision or complete inference requirements, so calling it "open source" is premature until those are published.

Reactions also raised concerns about inference economics, weaker results on some interactive tasks, benchmark presentation and allegations of undisclosed paid promotion. Latent Space notes that these are separate issues from the model's measured capability. No smaller Flash model was announced, though some users speculate that the "Pro" name implies one.