Kimi K3.1 Pro: Moonshot opens up a 2.8 trillion-parameter MoE with a one-million-token context
🔎 Moonshot delivers on its promise — and it was a big one
On November 4, 2026, Moonshot AI delivered exactly what it had announced: the Kimi K3.1 Pro weights are freely downloadable on Hugging Face, in a public, non-gated repository. No form, no waitlist, no identity verification clause. You click, you download.
The numbers are dizzying: a MoE with 2.8 trillion total parameters, of which roughly 104 billion are activated per token, a context window of 1,048,576 tokens, and a native vision encoder. While OpenAI and Anthropic jealously guard their closed weights, Moonshot is releasing a frontier-class model into the wild.
But the story doesn't end with the press release. Between benchmark claims to be taken with a grain of salt, a license with a 700-million-user clause, and 1.56 TB of data to download, this launch deserves a careful read. That's exactly what we're doing here.
The Essentials
- Architecture: MoE with 2.8 trillion total parameters, ~104 billion activated per token, 896 experts (16 selected + 2 shared per token), 93 layers — according to the official model sheet (November 2026).
- Context: 1,048,576 tokens, made cheaper by switching from full-attention layers to Gated DSA — the only architectural change compared to Kimi K3.
- Native Multimodal: MoonViT-V2 vision encoder with 401 M parameters, integrated into the model.
- Open Weights: non-gated repository created on November 2nd, 96 safetensors shards, 1.56 TB total, in MXFP4 precision with MXFP8 activations (quantization-aware training).
- Inference recommended by Moonshot: vLLM, SGLang, or TokenSpeed.
- Benchmarks: Moonshot claims to beat Opus 5.5 and GPT-6 Astra on paper; independent tests from MindStudio tell a "more complicated, and cheaper" story.
- License: free commercial use for up to 700 million monthly active users; beyond that, a separate agreement with Moonshot is mandatory.
Recommended Tools
| Tool | Primary Use | Price (November 2026) | Ideal for |
|---|---|---|---|
| vLLM | MoE Production Serving | Open source (actual cost = GPU) | Teams with a datacenter cluster |
| SGLang | High-concurrency Inference | Open source | Long context serving |
| TokenSpeed | Simplified Deployment | Open source | Testing and prototypes |
| Hugging Face | Weight Download (1.56 TB) | Free | Retrieving the 96 safetensors shards |
A 2.8 trillion MoE, but only 104 billion lit up at a time
Kimi K3.1 Pro is not a giant dense model: it's an extremely sparse mixture of experts. Of the 2.8 trillion total parameters, only ~104 billion are activated per token. Concretely, each query picks 16 experts out of 896, to which 2 shared experts are added — a classic architecture in very large MoEs that stabilizes training.
The whole thing rests on 93 layers, with a hybrid attention that makes up the model's entire appeal: 69 layers in Kimi Delta Attention (KDA) and 24 layers in Gated DSA. We'll come back to this below, but keep in mind that this hybridization is the key to the one-million-token context.
On the multimodal side, Moonshot isn't pretending: the MoonViT-V2 vision encoder (401 M parameters) is natively integrated. It's not a module grafted on after the fact, it's a building block of the model from the start — an approach in line with what the Kimi lineage has been doing since K3, which was also natively multimodal with a 1 M-token context according to the K3 model card.
The K3 lineage, in fast-forward
K3.1 Pro is a post-training of Kimi K3, Moonshot's 2.8 T base model, with large-scale joint reinforcement learning (RL) across hundreds of domains. Moonshot had already shown its cadence with Kimi K2.7-Code, the open-source 1T-parameter coding model that cuts 30% of reasoning tokens and beats Opus at tool use. The strategy is consistent: a massive base, specialized variants, and systematic open-sourcing of the weights.
Why 104 Billion Active Parameters Change the Economic Equation
A sparse MoE is a trade-off between computation and memory. The computation per token remains that of a ~100 billion parameter model—reasonable on a modern cluster. Memory, however, must host the 2.8 trillion complete parameters. This is where the precision chosen by Moonshot becomes decisive.
The weights are published in MXFP4 with MXFP8 activations, a precision derived from quantization-aware training. Translation: this is not a compressed version applied afterward; it is the precision with which Moonshot serves the model itself. The result: the entire model fits within 1.56 TB distributed across 96 safetensors shards.
For deployment teams, this changes two things. First, the cost per token for self-hosted inference becomes competitive against closed APIs—this is precisely what independent tests suggest, describing a "cheaper" model with comparable quality across several use cases. Second, the barrier to entry remains high: it requires hundreds of gigabytes of HBM memory, typically several datacenter GPU nodes. This is not a model you run on your workstation.
This is Moonshot's true bet: making inference economically viable at scale, not accessible to the individual. To achieve this, you need a cluster—and for now, few open weights players are in this league.
One million tokens of context, made viable by Gated DSA
The 1,048,576-token context window isn't just a marketing figure: it's the only architectural change from K3 that makes it economically sustainable. The 24 layers that were full-attention on K3 switch to Gated DSA, a sparse attention mechanism that drastically reduces the cost of long context.
Why does this matter? Because at 1M tokens, quadratic full attention becomes the dominant expense item. A model that keeps full attention on all of its layers burns GPU on long contexts; a model that hybridizes KDA (69 layers) and Gated DSA (24 layers) amortizes that expense. Moonshot isn't claiming a research breakthrough here, but an engineering trade-off: same architecture, cheaper long context.
As for use cases, this is the kind of model built for agents that swallow entire codebases, legal corpora, or weeks of conversation history. If you're exploring this territory, our guide to the best LLMs for AI agents details what to expect from a model for long-haul tool use.
Benchmarks: Moonshot's claims put to the test, independently
Moonshot claims that K3.1 Pro beats GPT-6 Astra and Claude Opus 5.5 on its internal benchmarks, including Terminal-Bench 4.0. We covered the numbers in detail in our article on Moonshot's victory on Terminal-Bench 4.0 and the weight release on November 4.
But independent testing tells a different story. The analysis by MindStudio (November 2026) is blunt: the claims of "beating Opus 5.5 and GPT-6 Astra" are true on paper, but evaluations conducted outside of Moonshot describe a model that is "more complicated, and cheaper." In other words: not a model that crushes closed-source competition across the board, but a model that delivers 90-95% of frontier performance at a fraction of the inference cost.
Our take: that's already huge. K3.1 Pro's real competitor isn't GPT-6 Astra, it's the price-to-performance ratio of closed APIs. On that front, an open 104B-active MoE is formidable — especially for companies that want to keep control of their inference stack.
One point of uncertainty to flag: Moonshot mentions a "K3.1 Pro Highspeed" version, presented as 2 to 3 times faster. No Highspeed checkpoint appears on Hugging Face at this stage. Separate checkpoint or just a serving mode? Nobody knows. We'll keep you posted.
Licensing: the real trap isn't the one you think
The "Kimi K3.1 Pro" license is identical, word for word, to that of K3. And it contains a clause that raises an eyebrow at first glance: any usage exceeding 700 million monthly active users requires a separate agreement with Moonshot.
Let's be clear: for 99.99% of companies, this clause is cosmetic. If you're building a SaaS, an agent, or a vertical product, you'll never come close to 700 million MAU — a threshold explicitly aimed at hyperscalers and consumer giants. It's an insurance policy against the scenario where a player the size of Meta or Google embeds the model into its product at planetary scale without negotiating.
The real trap lies elsewhere, and it's more discreet: it's the absence of a confirmed Highspeed checkpoint. If you size your infrastructure based on a promise of 2-3x higher speed that doesn't yet exist in the form of downloadable weights, you're taking an architectural risk. Our advice: size for standard K3.1 Pro, and treat Highspeed as a future bonus.
Another subtlety: the license is permissive for standard commercial use, but it is not a classic academic license like Apache or MIT. Have your legal team review the text if your use case falls outside the standard — redistribution, sublicensing, use in direct competition with Moonshot products.
Open weights: China has clearly taken the lead in the race
Kimi K3.1 Pro doesn't come out of nowhere. It's part of a sequence in which Chinese labs have taken control of the open weights narrative. Moonshot raised $2 billion in quick succession, as we detailed in our article on the $2 billion funding round and Kimi K2.6's open-weight dominance.
Competitors aren't standing still. Tencent has opened up HY4, a 770-billion-parameter model with native vision and computer use, whose release we covered in our article on Tencent HY4 in open weights. NVIDIA, for its part, is pushing Nemotron 3.5 Super Lightning, hybrid Mamba-Transformer MoEs that we analyze in our article on Nemotron 3.5 Super Lightning.
In our monthly comparison of the best LLMs, we see it every month: DeepSeek V4 Pro, GLM-5.1, and the Kimi models occupy the top of the open weights leaderboard, with scores that come close on the heels of closed models. What all these players have in common: they publish their weights, while leading American labs no longer do.
This is the most interesting reversal of 2026. Two years ago, open weights was a catch-up movement. Today, it's the innovation frontier — and it's mostly Chinese.
Who can actually run 1.56 TB?
Short answer: not you, not me, unless your company rents datacenter GPU nodes. With 1.56 TB of weights in MXFP4, you need HBM memory in industrial quantities — typically several 8x high-end GPU nodes. The model is designed for cloud serving, not for local use.
The process, on the other hand, is disarmingly simple. The repository is non-gated: no Hugging Face validation, no license agreement to sign online. You download the 96 shards, you spin up a vLLM or SGLang serving setup, and you're in production. Moonshot explicitly recommends vLLM, SGLang, or TokenSpeed as inference engines.
An important caveat for tinkerers: don't expect to re-quantize the weights to fp8 or int4 "to save space". The MXFP4/MXFP8 precision is native, the result of quantization-aware training. It's the reference precision, the one Moonshot itself serves. Any further re-quantization will start from an already-quantized model, with the quality degradation that implies.
And if your use case doesn't justify a cluster? Be realistic: smaller open models cover 80% of needs. Our guide to the best LLMs to run locally and our tutorial on installing a local LLM with Ollama or LM Studio cover the realistic options for a single-GPU machine.
One last practical point: if you're building a product on top of K3.1 Pro, separate the layers. Inference runs on your GPU cluster; the frontend, API, and business logic can run on standard web hosting like Hostinger — no need to pay for datacenter GPUs to serve HTML.
❌ Common Mistakes
Mistake 1: Confusing total parameters with active parameters
"2.8 trillion" does not mean that every request mobilizes 2.8 trillion parameters. Only ~104 billion are activated per token. Confusing the two leads to overestimating inference cost per request — and to dismissing a model that is actually competitive. Always read both figures.
Mistake 2: Sizing your infrastructure around the Highspeed version
The "2-3x faster" version has no confirmed repository on Hugging Face to date. If you architect your serving around that promise, you're taking an unnecessary technical risk. Size for the standard checkpoint, ship, and migrate if Highspeed actually comes out.
Mistake 3: Re-quantizing the MXFP4 weights "to save money"
The weights are already in MXFP4 with MXFP8 activations, a native precision trained in a quantization-aware manner. Re-quantizing an already-quantized model degrades quality with no real gain. If 1.56 TB doesn't fit within your GPU budget, pick a smaller model — not a mutilated version of this one.
Mistake 4: Ignoring the 700 million user clause
It's not a trap for you, but it is a real contractual commitment. If your product follows a hypergrowth, consumer-style trajectory, have your legal team run the numbers on that scenario now. Negotiating an agreement with Moonshot the day you cross the threshold is far harder than anticipating it.
❓ Frequently Asked Questions
Is Kimi K3.1 Pro really free?
Yes, for download and standard commercial use: the weights are openly available on a non-gated repository. The real cost is the infrastructure — 1.56 TB of MXFP4 weights require a datacenter GPU cluster. The only license restriction concerns uses beyond 700 million monthly active users, which require a separate agreement with Moonshot.
Can you run it on a local PC?
No, not as-is. With 1.56 TB of weights, even a high-end multi-GPU workstation falls far short. This model targets datacenter clusters served by vLLM, SGLang, or TokenSpeed. For local use, look at models with 7 to 70 billion parameters: our selection of the best local LLMs will point you toward realistic options.
Does K3.1 Pro really beat GPT-6 Astra?
On the benchmarks published by Moonshot, yes — notably Terminal-Bench 4.0. But MindStudio's independent tests (November 2026) paint a much more nuanced picture: the model is described as "more complicated, and cheaper" than the claims suggest. The real value lies in the performance-to-price ratio, not in universal frontier dominance.
What's different from Kimi K3?
K3.1 Pro is a post-training of K3 with large-scale joint RL across hundreds of domains. The only architectural change: the full-attention layers become Gated DSA, making long context cheaper. The license is identical word for word, and the MoE figures (2.8 T / 104 B active) are unchanged.
Does the 700 million user clause apply to me?
Almost certainly not. That threshold targets planet-scale platforms — think consumer giants. For a SaaS, a business agent, or a vertical product, even resounding success remains orders of magnitude below. It's a strategic protection clause for Moonshot, not a trap for the ecosystem.
✅ Conclusion
With Kimi K3.1 Pro, Moonshot proves that a lab can release 2.8T frontier weights while hiding nothing — neither about the architecture, nor about precision, nor about the license — and that cost-effective inference at scale is no longer the exclusive preserve of closed models. To find out where this MoE ranks against the competition this month, head over to our monthly comparison of the best LLMs.