Moonshot AI releases Kimi K3.1 Pro weights for self-hosting

Moonshot AI releases Kimi K3.1 Pro weights for self-hosting

Ondřej Barták
Ondřej Barták
Entrepreneur and Programmer
4. 11. 2026
2 minutes reading · 16 views
Listen to the article
Audio version of the article
Moonshot AI releases Kimi K3.1 Pro weights for self-hosting

Moonshot AI has made the complete Kimi K3.1 Pro weights available on Hugging Face, about two weeks after the model’s API launch. Teams can now examine, fine-tune and self-host the model, subject to its licensing terms and the hardware they have available.

Kimi K3 architecture and more extensive training

Kimi K3.1 Pro uses the same mixture-of-experts architecture as its predecessor, Kimi K3, activating only a subset of its parameters. It has approximately 2.8 trillion parameters in total, with about 104 billion active per token. The model accepts text and images and supports a context window of 1,048,576 tokens.

Moonshot says the model received substantially more intensive post-training, including joint reinforcement learning across hundreds of domains. The work aimed to improve tool use, planning and task completion. The company also says it removed cyber-related data from both training stages while keeping material for general-purpose programming.

Comparisons in agent benchmarks

In Moonshot’s published results, Kimi K3.1 Pro scored 65.7 on Terminal-Bench 4.0, compared with 66.4 for Claude Opus 5.5. On SWE-Marathon v1.1, Kimi scored 51.7 against Claude’s 46.3. The comparisons concern specific agent configurations and tasks.

Moonshot reported a score of 38.1 for Kimi on Terminal-Bench Science 0.1, versus 63.3 for Claude Opus 5.5 and 68.1 for GPT-6 Astra. In the company’s other comparisons, Kimi also trailed Claude Opus 5.5 on OSWorld V2.1 and BenchCAD, and the Claude and GPT models on ProgramBench and Humanity’s Last Exam without tools.

Artificial Analysis separately reported a 67.2% result for Kimi on Terminal-Bench 4.0. In that evaluation, Kimi ran mainly through Claude Code, while other models may have used different agent environments; the result reflects that setup as well as the model.

Self-hosting and API access

Self-hosting requires access to the complete set of weights. Those weights alone occupy approximately 5.6 TB at BF16 precision or 2.8 TB at 8-bit precision, excluding runtime overhead and the cache needed for long contexts.

At publication, Moonshot had not announced a separate license for Kimi K3.1 Pro. The earlier Kimi K3 uses a custom license that is not approved by the OSI; teams should check the repository for terms governing commercial use, modification and redistribution.

Alongside self-hosting, developers can access the model through Moonshot’s API as kimi-k3.1-pro or kimi-k3.1-pro-highspeed. Moonshot says Highspeed is functionally identical to the standard variant and two to three times faster. Both operate only in thinking mode; disabling it produces an HTTP 400 response.

The standard API costs $2 for one million input tokens and $10 for one million output tokens, a third less than Kimi K3’s stated rates; Highspeed is approximately twice as expensive per token.

Advertisement

Content created with help from UpTier.

SEO and GEO on autopilot. UpTier’s multi-agent systems write and optimize content for search engines and AI answers.

Discover UpTier ↗

Category:AI
Did you enjoy this article?
Discover more interesting posts on our blog
Back to blog

Related posts

Invisible Watermark in Grok-Generated Texts: How It Works and Who It AffectsInvisible Watermark in Grok-Generated Texts: How It Works and Who It Affects
Texts generated by Grok now carry a hidden marker that can be used to verify their origin. How does it work, who can disable it, and why isn’t it foolproof?
3 min read
10. 11. 2026
Anthropic releases Mythos 5.2 and Fable 5.2Anthropic releases Mythos 5.2 and Fable 5.2
Anthropic says its new models have improved at coding, reasoning and working with images. Fable 5.2 is also said to respond faster and supports enterprise zero-data-retention policies.
1 min read
10. 11. 2026
Meta releases Muse Ember, a lightweight model for AI agentsMeta releases Muse Ember, a lightweight model for AI agents
Muse Ember is designed for AI agents’ repetitive tasks and collaboration with a larger model. Meta says it can also process images, audio and video.
1 min read
9. 11. 2026
Přihlaste se k odběru našeho newsletteru
Zůstaňte informováni o nejnovějších příspěvcích, exkluzivních nabídkách, a aktualizacích.
CodedTrip

Operated by CodedTrip LLC, USA.

YouTube
TikTok