Kimi K3.1 Pro Open Weights: The Largest Free AI Model Ever

Published (Updated )

The quick answer
- Moonshot AI published the full weights of Kimi K3.1 Pro on Hugging Face on November 4, 2026, keeping a promise it made when the model first launched as an API and app on October 23, 2026.
- At 2.8 trillion total parameters, it is the largest open-weight AI model ever released, and it is also the strongest open model on the Artificial Analysis Intelligence Index.
- Anyone can download, inspect, fine-tune and run the model for free, but the licence has two carve-outs aimed at the biggest companies.
- The download is about 1.56 TB across 96 files, and running it takes serious hardware β roughly a node with eight or more top-end GPUs.
- The benchmark scores in this article are Moonshot's own reported numbers, not independently verified results, and they should be read that way.
What Moonshot actually released
Moonshot AI, the Chinese AI company behind the Kimi chatbot, put the weights of its flagship model Kimi K3.1 Pro on Hugging Face on November 4, 2026. The announcement went out on X at 16:00 UTC that day, and the public repository, moonshotai/Kimi-K3.1-Pro, was first seen live around 15:31 UTC.
This was not a surprise drop. When K3.1 Pro launched on October 23, 2026 as a hosted model β available on Kimi.com, the Kimi apps, Kimi Work, Kimi Code and the Kimi API β Moonshot said the open weights would follow. The November 4 release is the company keeping that promise, and it follows the same playbook it used with Kimi K3, which was open-sourced in July 2026.
So what does "open weights" actually mean here? It means the trained model files themselves are free to download. Researchers can look inside the model, developers can fine-tune it on their own data, and companies can run it on their own servers instead of paying per token through an API. It does not mean the training data or the full training code are public β those stay with Moonshot.
The release matters because of scale. Until now, the biggest open models sat far behind the frontier systems from Anthropic, OpenAI and Google. Kimi K3.1 Pro is the first open-weight release that Moonshot claims sits within a few points of the best proprietary models on independent-style leaderboards. That claim needs care β the numbers come from Moonshot itself β but even as a claim, it changes the conversation about how far open models have come.
How big is Kimi K3.1 Pro?
The headline number is 2.8 trillion total parameters. That makes it the largest open-weight model ever published, by a wide margin.
But the total number is only part of the story. Kimi K3.1 Pro is a Mixture-of-Experts model, which means only a fraction of its parameters are active for any single token. Here is how the architecture breaks down:
- Total parameters: 2.8 trillion
- Active parameters per token: 104 billion
- Experts: 896 in total, with 16 selected for each token
- Layers: 93
- Context window: 1,048,576 tokens β about 1 million tokens in one go
- Vision encoder: a 401-million-parameter MoonViT-V2 encoder, which gives the model native image and video understanding
- Inputs: text, images and video
- Reasoning: always on, with configurable effort levels so users can trade speed for deeper thinking
A million-token context window is one of the model's most practical features. In plain terms, it can hold roughly 750,000 words β several long books, a huge codebase, or months of documents β in a single conversation without forgetting the start. That is the kind of capacity that matters for real work: legal review, research, software projects and long-running agents that need to keep track of everything they have done.
The model was post-trained from Kimi K3 using large-scale joint reinforcement learning across hundreds of domains, according to Moonshot. That training approach is aimed at agentic work β tasks where the model plans, uses tools, writes code, checks its own results and keeps going until the job is done β rather than just answering a single question.
The benchmarks β and why they need a health warning
Every number in this section comes from Moonshot's own reporting. None of it has been independently verified yet, and independent labs will need time to run the model themselves. That matters, because vendors have every reason to present their best results. Read these scores as Moonshot's claims, not as settled fact.
With that warning in place, here is what Moonshot reports:
- Terminal-Bench 4.0: 65.7%, against 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra, according to Moonshot's own comparison table. Terminal-Bench tests how well a model can complete real software tasks in a terminal, so this is a test of agentic coding ability rather than chat skill.
- Artificial Analysis Intelligence Index: 55 points, ranking fourth overall, three points behind the best proprietary model, Claude Opus 5.5 at 58 points.
- Open-weight ranking: number one among 25 open-weight models on the Artificial Analysis index.
- The gap: Moonshot says the gap between the best open model and the best proprietary model has shrunk from 12 points to 3 points.
That last point is the one Moonshot wants everyone to remember. A year ago, open models lagged the frontier by a wide margin. If these numbers hold up under independent testing, the open frontier is now close enough that many companies could reasonably choose a free, self-hosted model over a paid API for serious work.
But there are reasons to wait for independent confirmation:
- Moonshot chose which benchmarks to highlight, and vendors always lead with their strongest results.
- Terminal-Bench scores can shift with the harness, the tooling and the number of attempts allowed.
- Real-world performance on a company's own tasks can differ a lot from leaderboard scores.
The honest position is this: Kimi K3.1 Pro looks like a genuinely strong model, and the open-weights community now has its best tool yet β but the claim that it is nearly as good as Claude Opus 5.5 is Moonshot's claim, and it still needs independent proof.
The licence: free, but not quite open source
This is where the release gets complicated, and it is the part most coverage skips over.
Kimi K3.1 Pro ships under the Kimi K3.1 Pro License. It reads like an MIT-style licence β download it, use it, fine-tune it, deploy it β but it is not a standard open-source licence, and it has two carve-outs aimed squarely at the largest players:
- Model-as-a-service businesses: if your business offers the model as a hosted service and earns more than $20 million in revenue over 12 months, you need a separate agreement with Moonshot before you can do that.
- Very large products: if your product has more than 100 million monthly users, or more than $20 million in monthly revenue, you must prominently display "Kimi K3.1-Pro" in the product.
For almost everyone else, the terms are generous:
- Anyone can download the weights for free.
- Anyone can inspect how the model works.
- Anyone can fine-tune it on their own data.
- Anyone can deploy it on their own hardware.
The practical effect is clear. Individual developers, researchers, startups and mid-size companies get a frontier-class model at no cost. The companies big enough to resell access to it as a service β the cloud providers and API platforms that would profit most from a free frontier model β have to negotiate with Moonshot first.
That is a deliberate business decision, and it is becoming a pattern. Several recent open-weight releases have used similar restrictions to stop the largest cloud companies from turning a free model into a paid service without the creator seeing any return. Whether that counts as "open" is a fair debate. What is not in doubt is that the weights are free for the vast majority of users.
Can you actually run it? The download and the hardware bill
Here is the part that separates the excitement from the reality. Kimi K3.1 Pro is free, but it is not light.
- Download size: about 1.56 TB, split across 96 safetensors files.
- Format: MXFP4 weights with MXFP8 activations, produced through quantization-aware training.
- GGUF version: none at release, so lightweight local running on a home PC is not available yet.
- Serving support: deployment recipes for vLLM (version 0.34.0 and above) and SGLang, plus TokenSpeed.
- Hardware: roughly a node with eight or more B300 or GB300-level GPUs, the current top tier of AI accelerator hardware.
To put that in perspective, 1.56 TB is more than most home broadband connections can pull down quickly, and eight top-end data-centre GPUs cost far more than a gaming PC β they are the kind of hardware that lives in a cloud data centre, not under a desk. A single one of these GPUs can cost tens of thousands of dollars.
That means the realistic audience for self-hosting is:
- Cloud providers and AI infrastructure companies.
- Large enterprises with their own GPU clusters.
- Research labs and universities with data-centre access.
- Specialist hosting companies that can rent access to the model.
For everyone else, the hosted API is the practical route β and that is where Moonshot still earns its money.

Hosted API pricing: what it costs if you don't self-host
For users who want to call the model through Moonshot's API instead of running it themselves, the public pricing is:
- Input tokens: $2 per million tokens
- Output tokens: $10 per million tokens
- Cached input tokens: $0.20 per million tokens
- API model id: kimi-k3.1-pro
At $2 per million input tokens and $10 per million output tokens, the hosted price sits in the same band as other frontier models, and the cheap cached-input rate of $0.20 per million tokens is aimed at long, repeated conversations where much of the context stays the same β exactly the kind of agentic work the model is built for.
Moonshot also offers a faster Highspeed variant that it claims runs two to three times faster than the standard version. That speed claim comes from Moonshot itself and has not been independently verified.
The model is live across Moonshot's own products: Kimi.com, the Kimi apps, Kimi Work, Kimi Code and the Kimi API. That matters because it means the weights release and the hosted product are the same model, not a cut-down version for the public.
Why this release matters beyond the numbers
The open-weights world has been waiting for a moment like this. For years, the pattern was fixed: the best models were closed, and open models were a step or two behind. Companies that wanted the best performance paid for an API. Companies that wanted control, privacy and no per-token bill settled for something weaker.
Kimi K3.1 Pro is the first release that credibly challenges that trade-off β if Moonshot's numbers hold up. A model that can be self-hosted, fine-tuned and run without sending data to a third party is a different proposition from an API, especially for:
- Regulated industries: banks, healthcare providers and government agencies in the US, UK and Europe that cannot send sensitive data to an outside API.
- Companies with huge token volumes: businesses running agents all day, where per-token API bills add up fast.
- Researchers: anyone who needs to study, test or modify the model itself rather than just call it.
- Startups: small teams that can now build on a frontier-class model without a large API budget.
There is also a wider context. Daily token consumption inside China has now passed 500 trillion tokens, according to a November 22 AI briefing β a sign of how fast AI use is growing there. Moonshot is one of the companies driving that growth, and releasing its flagship model openly is both a competitive move and a statement of confidence. It is betting that giving the model away will grow its ecosystem faster than keeping it locked up.
The release also puts pressure on Western labs. If a free model is within three points of the best paid model, the case for paying premium API prices gets harder to make β at least for the many tasks where the difference between first and fourth place does not matter much in practice.
What to watch next
Several things will decide whether this release lives up to the hype:
- Independent benchmarks: the first results from independent labs and from the Artificial Analysis team running the model themselves. If the scores hold, the story is confirmed. If they drop, the gap is wider than Moonshot claims.
- A GGUF release: a community or official GGUF version would let far more people run the model on smaller hardware. Nothing has been announced yet, and it is not clear whether the MXFP4 format will make that easy.
- Fine-tunes and derivatives: the open community usually produces tuned versions within weeks β smaller, faster or specialised variants. Watch Hugging Face for the first popular fine-tunes.
- Cloud availability: which cloud providers pick the model up, and what they charge for hosted access. That will decide how easy it is to use without owning a GPU cluster.
- Licence enforcement: whether Moonshot actually enforces the $20 million revenue and 100-million-user thresholds, and how the big cloud companies respond.
- The next frontier release: every open release pushes the closed labs to respond. The gap between open and closed models is now the number to watch in AI.

The bottom line
- Kimi K3.1 Pro is the largest open-weight AI model ever published, and by Moonshot's own numbers it is the strongest open model yet.
- The weights are free to download from Hugging Face, and the licence is generous for everyone except the very largest hosting businesses.
- Running it yourself takes about 1.56 TB of downloads and serious data-centre hardware, so most people will still use the hosted API at $2 per million input tokens and $10 per million output tokens.
- The benchmark claims are Moonshot's own and still need independent verification β but even if they come down a little, this is a major step for open AI.
- The release is from November 4, 2026, but it remains the biggest open-weights story of the month, and search interest in the model is still climbing this week.
For developers and companies in the US, UK and Europe, the practical advice is simple: try the hosted API first, watch the independent benchmarks, and only plan a self-hosted deployment once you know the model performs on your own tasks.
FAQs
What is Kimi K3.1 Pro?
Kimi K3.1 Pro is Moonshot AI's flagship AI model, released as an API and app on October 23, 2026. Its full weights were published openly on Hugging Face on November 4, 2026. It is a Mixture-of-Experts model with 2.8 trillion total parameters and a 1-million-token context window, and it can handle text, images and video.
Is Kimi K3.1 Pro really free to use?
Yes, for most people. Anyone can download, inspect, fine-tune and deploy the weights for free. The licence has two carve-outs for the largest players: model-as-a-service businesses earning over $20 million in 12 months need a separate agreement with Moonshot, and products with over 100 million monthly users or over $20 million in monthly revenue must display "Kimi K3.1-Pro" prominently.
How big is the Kimi K3.1 Pro download?
About 1.56 TB, split across 96 safetensors files. The weights use the MXFP4 format. There was no GGUF version at release, so running it on a home PC is not practical yet, and you need roughly a node with eight or more B300 or GB300-level GPUs to serve it.
How good is Kimi K3.1 Pro compared to Claude Opus 5.5?
According to Moonshot's own reported numbers, Kimi K3.1 Pro scores 65.7% on Terminal-Bench 4.0 against 66.4% for Claude Opus 5.5, and 55 on the Artificial Analysis Intelligence Index against 58 for Claude Opus 5.5. These are Moonshot's claims, not independently verified results, so treat them as the vendor's numbers until independent labs confirm them.
What does the Kimi K3.1 Pro API cost?
The hosted API costs $2 per million input tokens, $10 per million output tokens, and $0.20 per million cached input tokens. The API model id is kimi-k3.1-pro. A faster Highspeed variant is also available, which Moonshot claims runs two to three times faster.
Can I run Kimi K3.1 Pro on my own computer?
Not realistically on a normal home PC. The model needs about 1.56 TB of storage and data-centre-grade GPU hardware, and there is no lightweight GGUF version yet. Most users should use the hosted API or wait for cloud providers and the community to offer smaller, easier-to-run versions.
Was this article helpful?

Written By
TechorCoffeeContributing Writer
TechorCoffee is a technology reporter and software creator writing reviews and tutorials for TechOrCoffee.
Comments (0)
Leave a Comment
Read Next

Tue Oct 13 2026
Gemini 4 Argon: Everything You Need to Know
Gemini 4 Argon is Googleβs new flagship AI model. See its benchmarks vs GPT-6 Astra and Claude, pricing, 1M-token output, safety features and how to get access.

Fri Oct 30 2026
DeepSeek V4.1 Flash 1017: Cheap Coding AI Explained
DeepSeek V4.1 Flash 1017 explained for developers: 77.0 DeepSWE score, faster shorter outputs, Search and Fetch APIs, MIT open weights, and off-peak pricing.

Fri Nov 06 2026
AI Hackers Target Banks at Machine Speed, CERT-In Warns
CERT-In warns 'Crafty Falcon' is using AI tools to breach Indian lenders, fintech and crypto firms at machine speed. What the advisory says.