Moonshot AI Releases Kimi K3.1 Pro Open Weights: What Developers Need to Know

On November 4, 2026, Moonshot AI fulfilled its late-October roadmap commitment by releasing the open weights for its flagship model, Kimi K3.1 Pro. Published directly to an un-gated repository on Hugging Face, the release gives AI engineers, researchers, and enterprise teams direct access to a frontier-scale Mixture-of-Experts (MoE) model built for extended reasoning, multimodal input, and million-token workloads.

The Kimi K3.1 Pro open weights release arrives two weeks after Moonshot first launched the model across its consumer applications and developer API on October 23, 2026. In that initial technical announcement, the company stated that full model weights would be made publicly available by November 4. Following through on that timeline puts one of the largest open-weight models of 2026 directly into the hands of teams building custom coding assistants, multi-step agents, and private enterprise pipelines.

What Shipped in the Kimi K3.1 Pro Open Weights Release

According to the official Hugging Face model card and Moonshot AI’s release notes, the moonshotai/Kimi-K3.1-Pro repository provides the complete weights for the flagship K3.1 Pro checkpoint without an access gate or manual approval queue. Third-party repository tracking by ArtificialWatch on November 4, 2026 confirmed that the upload spans 96 safetensors shards totaling approximately 1.56 TB of model data.

At the core of the release is a sparse Mixture-of-Experts architecture designed to balance massive total capacity with manageable per-token compute:

• Total and Activated Parameters: Kimi K3.1 Pro contains 2.8 trillion total parameters, with 104 billion parameters activated per token.
• Expert Routing: Across 93 layers (including 1 dense layer), the model routes tokens across 896 routed experts, selecting 16 active experts per token alongside 2 shared experts.
• Attention Mechanism: The model pairs 69 layers of KDA with 24 layers of Gated DeepSeek Sparse Attention (Gated DSA). Moving from the full-attention Gated MLA used in Kimi K3 to Gated DeepSeek Sparse Attention reduces the computational overhead of long-context processing.
• Context Length and Vocabulary: The checkpoint supports a native context window of 1,048,576 tokens and uses a 160K vocabulary.
• Quantization-Aware Training: Rather than relying solely on post-training compression, Kimi K3.1 Pro uses MXFP4 weights and MXFP8 activations produced via quantization-aware training (QAT).
• Multimodal Vision Support: The model processes both text and images through an integrated 401-million-parameter MoonViT-V2 vision encoder.

One important boundary for developers to note is what did not ship in the Hugging Face upload. When Moonshot launched K3.1 Pro in late October, it also introduced a kimi-k3.1-pro-highspeed API tier, which the company describes as delivering two to three times faster output generation. As of November 5, 2026, Moonshot has only published the standard Kimi K3.1 Pro weights on Hugging Face; no separate Highspeed checkpoint has been released.

Why 1M Context and Open Weights Matter for Agent and Coding Workflows

Combining a 1,048,576-token context window with open weights addresses two of the biggest bottlenecks in production software engineering and autonomous agent workflows: context fragmentation and data governance. In complex coding tasks or multi-step agent loops, models routinely ingest entire repositories, long dependency trees, build logs, and dozens of sequential tool-call outputs. Understanding what a token is in AI makes the scale clear—a full million-token window lets an agent hold massive codebases and prolonged execution traces in working memory without aggressive summarization or brittle retrieval chunking.

Pairing that context capacity with open weights means engineering organizations no longer have to choose between long-horizon reasoning and strict infrastructure control. Teams working with proprietary source code, regulated financial records, or internal infrastructure telemetry can deploy the model inside their own virtual private cloud or on-premises cluster, fine-tune it on domain-specific tool schemas, and lock in deterministic behavior without worrying about silent upstream API deprecations.

Agent Arena Claims and Moonshot’s Reported Benchmarks

The arrival of the Kimi K3.1 Pro open weights has drawn immediate attention from independent evaluation platforms, particularly around autonomous agent execution.

On November 4, 2026, the LMSYS Arena team (@arena) posted an update on X stating that Kimi K3.1 Pro had taken the #1 spot among open-weight models on the Agent Arena leaderboard. According to LMSYS Arena’s statement—which is based on more than 17,500 real-world, long-horizon agent sessions—Kimi K3.1 Pro (Max) achieved a +5.1% net improvement, ranking #1 among open models and #13 overall across both proprietary and open systems. Arena also stated that the model ranks #1 among open-weight models across all three tracked tracks: Work, Code, and Chat, improving by 4.2 percentage points over the previous Kimi K3 (Max) release.

Alongside third-party leaderboard tracking, Moonshot AI published its own internal benchmark evaluations in its technical blog post:

• Terminal-Bench 4.0: Evaluated using the Claude Code harness, Moonshot reports that Kimi K3.1 Pro scored 65.7%, compared to the blog’s reported figures of 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra.
• BenchCAD with Tools: In tool-assisted CAD tasks, Moonshot reports a score of 92.7%.

Because vendor-reported benchmarks and community leaderboards each reflect specific harnesses and prompt distributions, engineering teams evaluating open-weight models in 2026 should treat both figures as directional signals rather than universal guarantees. Testing the model against your own repository tasks and failure modes remains essential when deciding how to choose an AI model for production.

Kimi K3.1-Pro License Caveats

While the weights are publicly downloadable on Hugging Face without a registration wall, Kimi K3.1 Pro is not released under a pure, unrestricted MIT or Apache 2.0 license. Instead, Moonshot uses the custom Kimi K3.1-Pro License, which starts from a permissive MIT-style baseline and attaches specific commercial and attribution thresholds.

The following summary highlights the main operational clauses for developer teams, though it is provided for informational purposes only and is not legal advice:

1. Copyright and License Notice: Any distribution or derivative work must retain the original copyright notice and the Kimi K3.1-Pro License text.
2. Model-as-a-Service (MaaS) Revenue Cap: If you offer Kimi K3.1 Pro (or derivatives) externally as a hosted Model-as-a-Service—such as a commercial inference API—and the combined gross revenue of the licensee and its affiliates exceeds USD 20 million over any rolling 12-month period, you must execute a separate commercial agreement with Moonshot before commercializing.
3. High-Scale UI Attribution: If a product or service built with the model exceeds 100 million monthly active users (MAU) or generates more than USD 20 million in monthly revenue, the product must prominently display “Kimi K3.1-Pro” in its user interface.
4. Internal Use and Official Channels: Internal company use, research deployments, and applications accessed through Moonshot’s official products or certified inference partners are exempt from the MaaS and high-scale attribution restrictions.

Organizations that expect to cross these commercial thresholds or build public inference endpoints can contact Moonshot directly at [email protected].

How to Access Kimi K3.1 Pro via Hugging Face or API

Developers have two main paths for testing and deploying Kimi K3.1 Pro depending on their infrastructure budget and latency goals.

Self-Hosting From Hugging Face

Teams with dedicated multi-node GPU clusters can pull the 96 safetensors shards (~1.56 TB on disk) directly from the moonshotai/Kimi-K3.1-Pro repository on Hugging Face. The official model card recommends serving the checkpoint using vLLM, SGLang, or TokenSpeed, all of which support the model’s MoE routing, Gated DeepSeek Sparse Attention layers, and MXFP4/MXFP8 quantization format.

Managed API and Developer Tools

For teams that prefer to avoid the operational complexity of hosting a 2.8-trillion-parameter model—or those running an AI API pricing comparison before committing to hardware—Moonshot offers hosted access through platform.kimi.ai under the model IDs kimi-k3.1-pro and kimi-k3.1-pro-highspeed, as well as inside Kimi.ai, Kimi Work, and Kimi Code.

The official platform supports both OpenAI-compatible and Anthropic-compatible API formats. When calling kimi-k3.1-pro for multi-step agent or coding tasks, the model card highlights two critical implementation details:

• Configurable Reasoning Effort: You can control test-time compute via the reasoning_effort parameter, which accepts low, high, or max (defaulting to max).
• Multi-Turn State Preservation: In multi-turn conversations and tool-use loops, your application must pass back the complete assistant message from prior turns—including both reasoning_content and tool_calls—so the model maintains coherent reasoning state across steps.

FAQ

Where can I download the Kimi K3.1 Pro open weights?

You can download the weights directly from the un-gated moonshotai/Kimi-K3.1-Pro repository on Hugging Face. The release consists of 96 safetensors shards totaling roughly 1.56 TB.

Did Moonshot release open weights for Kimi K3.1 Pro Highspeed?

No. As of November 5, 2026, only the standard Kimi K3.1 Pro checkpoint has been published on Hugging Face. The kimi-k3.1-pro-highspeed variant remains available through Moonshot’s hosted API and official applications.

Is Kimi K3.1 Pro free for commercial use?

For most internal deployments and small-to-midsize products, yes, provided you keep the license notice intact. However, if you sell hosted MaaS/API access and your organization earns over USD 20 million in rolling 12-month revenue, you need a separate commercial license from Moonshot. Products exceeding 100 million MAU or USD 20 million in monthly revenue must also display “Kimi K3.1-Pro” in their UI. Always consult legal counsel for your specific use case.

What inference engines support Kimi K3.1 Pro?

Moonshot’s Hugging Face model card recommends deploying Kimi K3.1 Pro with vLLM, SGLang, or TokenSpeed.

How do I configure reasoning depth in the Kimi K3.1 Pro API?

The API exposes a reasoning_effort parameter that can be set to low, high, or max (with max as the default). For multi-turn tool use, make sure to return the full assistant message—including reasoning_content and tool_calls—in subsequent requests.

Sources

• Moonshot AI Official X Announcement: Releasing the Model Weights of Kimi K3.1 Pro (Nov 4, 2026)
• Kimi K3.1 Pro Model Card and License on Hugging Face (moonshotai/Kimi-K3.1-Pro)
• Moonshot AI Technical Blog: Kimi K3.1 Pro
• LMSYS Arena (@arena) Agent Arena Statement on X (Nov 4, 2026)
• ArtificialWatch: Kimi K3.1 Pro Open Weights Release Report (Nov 4, 2026)