Kimi K3.1 Pro (scx-ai-gp)

Moonshot AI introduced Kimi K3.1 Pro as the first model in its new K3.1 series, describing it as its most capable 3T-class open-weight model to date with a one-million-token context window. A companion variant, Kimi K3.1 Pro Highspeed, reportedly delivers two to three times the speed with identical capabilities, and is exposed through the Kimi API as kimi-k3.1-pro and kimi-k3.1-pro-highspeed, with zero data retention available for API customers. K3.1 Pro is a post-trained version of Kimi K3 built with substantially larger post-training compute, using large-scale joint reinforcement learning across hundreds of downstream domains after optimizing the RL framework, long-context inference, and rollout scheduling. Moonshot reports 65.7% on Terminal-Bench 4.0, 61.2% on Kimi Code Bench 3.0 Long-horizon versus K3's 55.95%, and 92.7% with tools on BenchCAD.

Third-party aggregator BenchmarkList tracks Kimi K3.1 Pro across 35 benchmark rows, placing it in a field that includes Claude Opus 5.5, Gemini 4 Carbon, MiMo V2.6 Pro, and DeepSeek V4.1 Flash. Listed scores include a GDPval-AA Elo of 1846 at the 99th percentile, an AA-Briefcase Elo of 1678, a BrowseComp score of 92.1% in max thinking mode, and 73.3% partial on AutomationBench-AA. The page also records per-model pricing per one million tokens, listing Kimi K3.1 Pro at $2.00 in and $10.00 out, alongside competitor rates, and marks each evaluation row as self-reported and dated 2026-10-23. Because the entries are vendor-reported figures in a marketplace-style leaderboard, the comparisons should be treated as indicative rather than independently reproduced results.

Microsoft Foundry's model catalog lists Kimi-K3.1-Pro as a Moonshot AI model available directly from Azure, describing it as an open-weight, natively multimodal agentic model with a 1-million-token context window and gains in coding, 3D creation, financial research, and factual accuracy. It notes the model is post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of downstream domains. The catalog entry highlights long-horizon coding with minimal human oversight across massive repositories and terminal tools, covering GPU kernel optimization, compiler work, vision-in-the-loop game development, CAD, and chip design, plus agentic knowledge work producing interactive visualizations, widgets, and dashboards. It also frames the model for unified Azure billing, governance, and pay-as-you-go or reserved PTU deployment.

Artificial Analysis independently evaluates Kimi K3.1 Pro, scoring it 46 on its Intelligence Index, well above the median of 23 among comparable open-weight reasoning models. It lists the model as open weights with a 1M-token context, 2.8T total and 104B active parameters, text and image input, and a Kimi K3.1-Pro Licence with weights hosted on Hugging Face. On cost, the page reports $2.00 per million input tokens and $10.00 per million output tokens with a 90% cache discount, calling the model expensive relative to peers at roughly $3.68 per Intelligence Index task. Performance measurements include 174.78s time to first token, 33 tok/s output speed, about 36.2 minutes per Index task versus a 14.2-minute median, and a verbose 113M generated tokens.

Moonshot's official Kimi API platform documentation introduces Kimi K3.1 Pro and the Kimi K3.1 Pro Highspeed variant, which delivers 2-3x the speed of K3.1 Pro for more responsive interactions. Both are described as the first models in the new K3.1 series, available with a 1-million-token context window, and the docs position the release as the provider's most capable 3T-class open-weight model to date. The quickstart guidance frames K3.1 Pro as a post-trained version of Kimi K3 using substantially larger post-training compute and large-scale joint reinforcement learning across hundreds of downstream domains, with the RL framework, long-context inference, and rollout scheduling reworked. It reports Terminal-Bench 4.0 at 65.7% and the internal Kimi Code Bench 3.0 Long-horizon at 61.2%, up from K3's 55.95%.

The Hugging Face model card for moonshotai/Kimi-K3.1-Pro details the open-weight release as a natively multimodal agentic model with a Mixture-of-Experts architecture: 2.8T total parameters, 104B activated per token, 93 layers, 896 experts with 16 selected plus 2 shared, 160K vocabulary, and a 1,048,576-token context length with text and image input. Weights are quantized in MXFP4 with MXFP8 activations under quantization-aware training. Notably, compared with Kimi K3 the release replaces the full-attention Gated MLA layers with Gated DeepSeek Sparse Attention (Gated DSA) layers for more efficient long-context modeling, alongside 69 KDA and 24 Gated DSA attention layers. The card also publishes creator-run evaluation tables where Terminal-Bench 4.0 is listed at 65.7% for Kimi K3.1 Pro against 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra.

Moonshot AI introduced Kimi K3.1 Pro, the first model in its new K3.1 series and its most capable 3T-class open-weight model to date, with a 1-million-token context window and substantial gains in coding, 3D creation, financial research, and other agentic work. It is a post-trained version of Kimi K3 built with substantially larger post-training compute, using large-scale joint reinforcement learning across hundreds of downstream domains. Moonshot says it optimized the RL framework, long-context inference, training efficiency, sandbox infrastructure, rollout scheduling, and environment quality checks so each increment of RL compute returns stable gains. On Terminal-Bench 4.0 the model scores 65.7%, ahead of GPT-6 Astra (58.2%) and just behind Claude Opus 5.5 (66.4%), while on the vendor-run Kimi Code Bench 3.0 Long-horizon it reaches 61.2% versus K3's 55.95%. A Highspeed variant delivers 2-3x the speed.