Moonshot Kimi K3.1 Pro Launches With 1M Context Window and Open Weights Promise

Moonshot AI released Kimi K3.1 Pro on October 23, 2026, as the first model in its K3.1 series. The post-trained upgrade to the July 2026 Kimi K3 targets coding and agentic workloads with a 1M-token context window and deliberate reductions in offensive cyber capabilities.

Moonshot Kimi K3.1 Pro Architecture and Availability

The model uses a 2.8T total parameter MoE design with 104B active parameters across 896 experts, activating 16 per token. It supports multimodal inputs including text and images, with some video capabilities reported. The system always reasons before responding. A Highspeed variant delivers 2–3 times faster inference at roughly double the price.

Access opened immediately through Kimi.ai, Kimi Work, Kimi Code, and the Kimi API under the identifiers kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Pricing stands at $2 per million input tokens and $10 per million output tokens, with cached inputs at $0.20. The Highspeed tier costs approximately twice as much. A free tier carries usage limits, while enterprise features remain disabled by default. Zero Data Retention is available for API users.

Moonshot promised full open weights release by November 4, 2026. The company positions the model as the strongest 3T-class open-weight option available once released. Developers can expect integration through platforms such as Vercel and OpenRouter after the weights appear on Hugging Face.

Benchmarks, Cyber Reductions, and Developer Implications

Moonshot reported Terminal-Bench 4.0 scores of 65.7 percent under maximum reasoning effort, close to Claude Opus 5.5 at 66.4 percent and ahead of GPT-6 Astra at 58.2 percent. SWE-Marathon reached 51.7 percent, leading listed models. The internal hallucination rate fell to 8.3 percent from K3’s 23.4 percent. BenchCAD with tools scored 92.7 percent, with strong results also noted in finance and long-horizon agent tasks.

Artificial Analysis placed the Intelligence Index at 55, fourth among 215 models and 11 points above K3’s 44. The model trails only Claude Opus 5.5 (58), Claude Sonnet 5.5 (56), and Muse Wonder Full (56). Agentic metrics rank near the top, including Terminal-Bench 4.0 near 67 percent and GDPval-AA at 1,846 Elo, up 312 points from K3. The Cyber Index reached 64 percent, the highest among openly available models and 23 points above K3.

Moonshot removed cyber-related data from mid-training and RL environments before open release. Internal xdev-bench-hard dropped to 81.0 percent from 90.5 percent on the unfiltered checkpoint, compared with K3 at 33.3 percent. The company stated the change preserves general coding strength while limiting extension of offensive cyber capabilities. Future iterations may restore defensive cyber functions gradually.

Context, Cost Advantages, and Release Timeline

Kimi K3.1 Pro builds on the original K3 through expanded post-training and RL compute across hundreds of domains. It competes with leading proprietary models on coding and agent tasks at lower token prices. The 1M context window supports long-document and multi-step agent workflows that smaller-context models handle less efficiently.

Enterprise teams gain practical value through lower costs than K3 ($3/$15) and the forthcoming open weights. Independent verification of benchmarks and broader adoption through API gateways are expected after November 4. Moonshot may adjust cyber capability levels in subsequent open models based on observed usage patterns.