Moonshot AI officially unveiled Kimi K3.1 Pro, an advanced large language model designed to rival Western flagship systems such as Anthropic’s Claude Opus 5.5 and OpenAI’s GPT-6 Astra.
The launch marks the Chinese artificial intelligence (AI) startup’s most capable 3-trillion-parameter-class system to date. Built upon July’s Kimi K3 architecture — which coordinates 896 individual neural networks while activating just 16 at a time to curb inference costs — K3.1 Pro also retains Kimi Delta Attention, a memory-efficient mechanism that reduces RAM usage during prompt processing.
Moonshot AI’s new model underwent a heavily upgraded post-training regimen supported by automated quality checks and expanded compute infrastructure.
Engineered specifically for agentic workflows, complex reasoning, and code execution, Kimi K3.1 Pro features a massive 1-million-token context window alongside native multimodal support for text, image, and video inputs. By default, Moonshot AI’s API operates the model with reasoning permanently engaged at maximum effort, emphasizing rigorous problem-solving over raw latency.
According to internal evaluations released by the company, K3.1 Pro posts highly competitive results across technical disciplines. On the Terminal-Bench 4.0 software coding benchmark, K3.1 Pro recorded a 65.7% success rate, trailing Claude Opus 5.5 (66.4%) by less than a point while comfortably surpassing GPT-6 Astra (58.2%). The model also outpaced both rivals on OfficeQA Pro v2, an industry test measuring complex financial document analysis.
The system demonstrated notable gains in programmatic 3-D modeling and animation. Rather than drawing geometry visually, K3.1 Pro generates CAD code to define spatial structures, scoring 92.7% on the BenchCAD benchmark, within four percentage points of leading competitors. Internal testing demonstrated its capacity to synthesize complex human motion simulations and generate executable 3-D video game code.
Despite these gains, the model trails significantly in cybersecurity assessments. Moonshot AI acknowledged the gap, stating engineers deliberately purged vulnerability analysis data from post-training datasets to mitigate safety risks before broader distribution.
The announcement arrives as enterprise developers pivot from conversational chatbots to autonomous agents.
While a 1-million-token context window allows deep integration with extensive codebases, analysts note that organizations must still conduct rigorous enterprise audits to verify retrieval accuracy across long contexts, enforce strict internal data boundaries, and evaluate output latency.
Moonshot AI has made Kimi K3.1 Pro immediately available alongside a lower-latency sibling, Kimi K3.1 Pro Highspeed, across Kimi.ai, Kimi Work, Kimi Code, its developer API, and third-party platforms including Vercel AI Gateway. Moonshot AI plans to release the model’s open weights to the public by Nov. 4.
