Moonshot launched Kimi K3.1 Pro, a multimodal mixture-of-experts model with 2.8 trillion total parameters and 104 billion active parameters, scoring 55 on the Artificial Analysis Intelligence Index. The model placed fourth behind Opus 5.5 at 58 and Sonnet 5.5 at 56, while beating Fable 5.1 on independent tests. Weights are promised for November 4, 2026, which matters because the open-versus-proprietary gap would narrow from 12 points to three.
Launch terms and promised open release
Moonshot describes K3.1 Pro as a post-trained version of Kimi K3 with substantially larger post-training compute and a 1-million-token context window. Advertised strengths include full-project coding, Blender and CAD work, financial modeling, and long-running agents. Standard and Highspeed versions launched across the website and app, Kimi Work, Kimi Code, and API, plus Kimi Hosted Agents with a 24/7 environment for long-running work and multi-agent collaboration. Highspeed is said to preserve intelligence while generating output two to three times faster at twice the standard price.
The model was built through large-scale joint reinforcement learning across hundreds of downstream domains, with optimization of the RL framework, long-context inference, training efficiency, sandbox infrastructure, rollout scheduling, and shared environment checks. Moonshot says each increment of RL compute delivered stable returns and consistent gains in every domain. The update resembles a minor version bump driven by much more post-training, similar to the earlier Z. ai 5.2-5.3 step. No layer layout, routing configuration, attention design, or training-compute budget was disclosed in the materials reviewed.
At launch the API was available while weights remained unpublished, so Artificial Analysis recorded them as unpublished and the precise term is soon-to-be open-weight rather than open source. The announcement gives November 4, 2026 as the date, following earlier rumors of a release tied to November 5 and reports of staged availability on kimi. com before kimi. ai. License terms, redistribution conditions, commercial-use limits, checkpoint precision, and full inference requirements are still unknown. No smaller Flash model was announced, despite speculation that the Pro label implies one later.
What this means for enterprise AI adoption
For companies running coding and agentic workloads, the independent results point to practical gains in execution rather than general knowledge. Artificial Analysis reports accuracy up from 48 percent to 51 percent and hallucination rate down from 53 percent to 44 percent on AA-Omniscience versus K3. Arena placed K3.1 Pro eighth in Code Arena WebDev with 1,692 points, up nine places and 37 points over K3, roughly matching Opus 5 Max. On cyber tasks it scored 64 percent despite reduced cyber training, the highest publicly accessible result and second overall.
Standard pricing is 2 dollars per million input tokens, 10 dollars per million output tokens, and 0.20 dollars per million cached input tokens, down one third from 3, 15, and 0.30 dollars for K3. That token discount does not guarantee cheaper completed work, because large agentic runs consume long context and repeated tool calls, and Highspeed doubles the rate. The 104 billion active figure also does not equal a 104B dense model to host: the full 2.8T expert pool implies about 5.6 TB at 16-bit or 1.4 TB at 4-bit before overhead, requiring substantial accelerator infrastructure.
The marker to watch is the actual weight publication around November 4 and the license that accompanies it. If downloadable checkpoints appear with commercial-use permission and third-party serving benchmarks confirm the 55-point level outside Moonshot infrastructure, the three-point frontier gap becomes an operational option for private deployment. If licensing restricts use or serving costs erase the token saving, the release will remain an API alternative to Opus and Fable rather than a shift in enterprise infrastructure.
