![Kimi K3.1 Pro Goes Open-Weight: 2.8T Parameters [2026]](https://shattered.io/wp-content/uploads/2026/11/moonshot-kimi-k3-1-pro-open-weights-2-8-trillion-parameters-2026-gen.webp)
Moonshot AI has released Kimi K3.1 Pro as an open-weight model, with the files hosted under moonshotai/Kimi-K3.1-Pro on Hugging Face, according to TechNode’s reporting and the model listing itself. The headline figures are huge: 2.8 trillion total parameters, 104 billion active per token, and a context window of 1,048,576 tokens.
The release lands in a busy stretch for open models. Our open-weight spec sheet tracks how these checkpoints stack up, and Kimi K3.1 Pro adds a new row with an unusually large number in the parameter column.
Here is what the sources do and do not establish as of November 05, 2026. The model card pages are dated November 2 through 5, so no single official release date is confirmed. No price appears in the material available to us. No attributable on-record quote from Moonshot AI surfaced either. This analysis sticks to the published specifications, works out what they imply with plain arithmetic, and labels every judgment call as ours.
The short version: Kimi K3.1 Pro is a sparse Mixture-of-Experts model that activates only a sliver of its weights for each token, ships in low-precision formats to keep the download manageable, and takes text and images as input. Whether it matters to your stack depends on a license we have not been able to read in full and on hardware most teams do not own.
Don't miss new tech stories on Google
Add Shattered once in the Google app and our stories show up in your news suggestions.
Kimi K3.1 Pro Specs at a Glance #
Start with the confirmed numbers. Everything in the table below comes from the Hugging Face model card and the coverage that repeats it. We added nothing and rounded nothing.
| Specification | Kimi K3.1 Pro |
|---|---|
| Developer | Moonshot AI |
| Model identifier | kimi-k3.1-pro |
| Hugging Face repository | moonshotai/Kimi-K3.1-Pro |
| Architecture | Mixture-of-Experts (MoE) |
| Total parameters | 2.8 trillion |
| Active parameters per token | 104 billion |
| Context length | 1,048,576 tokens |
| Experts | 896 |
| Experts selected per token | 16 |
| Vocabulary size | 160,000 |
| Vision encoder | MoonViT-V2, 401 million parameters |
| Numeric formats | MXFP4 weights, MXFP8 activations |
| Inputs | Text and images |
| Lineage | Post-trained from Kimi K3 |
| License | Kimi K3.1 Pro License |
| Repository size on Hugging Face | 1.56 TB |
Two details stand out right away. The expert count is very high, and the weights ship in 4-bit-class precision. Both choices point to a team optimizing for serving cost rather than for a model that fits on one box.
Where the Weights Live and What 1.56 TB Means #
The Hugging Face repository shows a storage size of 1.56 TB. That is the first number any would-be self-hoster should read twice. A download of that size ties up a serious chunk of disk, takes real time on a typical connection, and needs a plan for where the files will sit once they land.
The sources we have cover the model files and specs, not the recommended inference stack. Check the model card for instructions before you pull anything. If you only want to confirm that your storage can hold the repository, a quick check looks like this:
df -h /data
huggingface-cli download moonshotai/Kimi-K3.1-Pro --local-dir /data/kimi-k3.1-pro
A Sanity Check on the Footprint
The size lines up with the stated formats. A model with 2.8 trillion parameters stored at 4 bits each would take about 1.4 TB before any overhead. MX-style formats attach a shared scale factor to small blocks of values, which adds a little on top. The listed 1.56 TB sits just above the raw 4-bit figure, which fits a checkpoint that is mostly MXFP4. This is our arithmetic, not a Moonshot AI statement, and the sources do not break the repository down by component.
Why a 2.8 Trillion Parameter Model Activates Only 104 Billion #
A Mixture-of-Experts model splits its feed-forward layers into many specialist blocks and routes each token through only a few of them. The rest sit idle for that token. Compute per token tracks the active parameter count, while memory tracks the total.
For Kimi K3.1 Pro, 104 billion active out of 2.8 trillion total works out to about 3.7 percent. Per token, the model does roughly the arithmetic of a 104 billion parameter dense network. It still needs to keep all 2.8 trillion parameters reachable, because the router can send the next token anywhere.
That split explains who benefits. A provider serving thousands of concurrent requests can spread the full weight set across many accelerators and keep each one busy, so the low active count turns into cheap tokens. A single developer with one workstation gets the opposite deal. The memory bill arrives in full, and the compute savings do not rescue you.
896 Experts, 16 Per Token: The Routing Math #
The model card lists 896 experts with 16 selected for each token. That is a fine-grained design. Many small specialists, several of them consulted at once, give the router more combinations to choose from than a layout with a handful of large experts.
The table below turns the published figures into ratios. Every row is simple division on numbers from the fact sheet or from our earlier coverage. None of these ratios come from Moonshot AI.
| Derived figure | Inputs | Result |
|---|---|---|
| Share of parameters active per token | 104B / 2.8T | About 3.7% |
| Share of experts used per token | 16 / 896 | About 1.8% (1 in 56) |
| Context length as a power of two | 1,048,576 tokens | 2 to the 20th power |
| Repository bytes per total parameter | 1.56 TB / 2.8T | About 0.56 bytes |
| Vision encoder share of total parameters | 401M / 2.8T | About 0.014% |
| Total size vs. Tencent Hy4 headline | 2.8T / 770B | About 3.6x |
The gap between 3.7 percent of parameters and 1.8 percent of experts is worth a second look. Experts make up most, but not all, of the model. Attention layers, embeddings and any shared components run for every token, which pushes the active parameter share above the expert share. The sources do not say how Moonshot AI splits the 104 billion between shared and routed parts, so we leave that alone.
The 1,048,576-Token Context Window #
Moonshot AI lists a context length of 1,048,576 tokens, or 1 million in round terms. In practice that means entire code repositories, long contract bundles or hours of transcribed meetings can sit in a single prompt.
A nominal context length and a usable one are different things. Serving a million tokens costs memory for the key-value cache, and quality can sag on retrieval tasks deep inside a long prompt. The sources we have say nothing about the attention design or about long-context test results. Until independent evaluators publish numbers, treat the headline as a ceiling, not a guarantee.
Agent workloads make the figure more interesting than it first looks. A coding agent that reads files, runs tools and appends every result to its history burns through context quickly. Long windows let such an agent keep its full trail instead of summarizing and forgetting. That is a plausible reason Moonshot AI pairs the context size with the agentic label, though the company’s sources here do not spell it out.
MXFP4 Weights and MXFP8 Activations #
The model card lists MXFP4 weights and MXFP8 activations. Both belong to the microscaling family, where small groups of values share one scale factor so that very low bit widths keep usable accuracy. Weights at 4 bits shrink the checkpoint and the memory traffic. Activations at 8 bits keep the intermediate math a little more precise.
Shipping in these formats is a statement about target hardware. Accelerators with native support for low-precision formats can run such a checkpoint efficiently, while older chips may need conversion or emulation that eats the gains. The sources do not list supported hardware or inference engines, so buyers should confirm compatibility before ordering anything.
There is also a distribution upside. A 1.56 TB repository is large but not absurd, and a full 16-bit copy of 2.8 trillion parameters would run to roughly 5.6 TB by the same arithmetic. Releasing the low-precision version makes the open-weight promise practical for more organizations.
MoonViT-V2 and the Natively Multimodal Agentic Pitch #
Moonshot AI describes Kimi K3.1 Pro as a natively multimodal agentic model, with text and images as listed inputs. The vision side runs on MoonViT-V2, an encoder with 401 million parameters. Against the 2.8 trillion total, that encoder is a rounding error, roughly 0.014 percent.
A small encoder feeding a huge language backbone is a common pattern. The encoder turns pixels into tokens, and the big model does the reasoning. For developers, the practical payoff is screenshots, charts and scanned documents entering the same prompt as text. Agents that operate software through its interface benefit most, since they can read the screen directly.
The sources list text and images only. They do not mention audio or video input, so do not plan around either. They also do not describe how the model performs on any vision test, which leaves the word native as a design claim we cannot yet check.
Post-Trained From Kimi K3: What the Lineage Tells You #
The model card says Kimi K3.1 Pro was post-trained from Kimi K3. That wording matters. A point release built on an existing base usually changes behavior through additional training stages, not through a new architecture or a fresh pretraining run. Expect continuity in the tokenizer, the expert layout and the general knowledge base, with differences showing up in instruction following, tool use and reasoning habits.
For teams already running Kimi K3, that suggests an upgrade path with fewer surprises than a ground-up successor would bring. We cannot confirm it, because the sources do not document what changed between the two models. Moonshot AI has not given a changelog in the material we reviewed, so run your own regression tests before swapping anything in production.
The lineage also hints at cadence. Labs that post-train from a strong base can ship improvements quickly, and the 3.1 label suggests Moonshot AI intends to keep iterating on the K3 family instead of waiting for a K4.
The Kimi K3.1 Pro License Question #
The weights ship under something called the Kimi K3.1 Pro License. We could not verify its terms from the sources available, so we make no claims about what it allows. That gap is the most important open question in this release for any company that wants to build a product on the model.
The name alone tells us something. It is a bespoke license, not a standard one. Our coverage of the Tencent Hy4 release noted its Apache 2.0 terms, which come with well-understood commercial rights. A custom license can be just as permissive, or it can carry conditions that surprise a legal team. Open-weight and open-source are not synonyms.
What to Check Before You Build on It
- Whether commercial use is allowed, and under what conditions or thresholds.
- Whether you may redistribute fine-tuned or quantized derivatives.
- Whether the license requires attribution or a specific naming of derived models.
- Whether it limits certain use cases or deployment regions.
- Whether training other models on its outputs is permitted.
Read the license file in the repository itself, not a summary of it, including ours. Have counsel look at it if revenue depends on the answer.
How Kimi K3.1 Pro Compares With Other Open-Weight Releases #
Direct comparisons are tricky right now. Independent benchmark results for Kimi K3.1 Pro were not in the sources, and parameter counts say little about quality on their own. What we can compare is scale, licensing posture and release status, using figures from our own reporting. Treat the table as context, not a leaderboard.
| Model or release | Headline figure | License or status | Where we covered it |
|---|---|---|---|
| Kimi K3.1 Pro (Moonshot AI) | 2.8T total, 104B active, 1,048,576-token context | Kimi K3.1 Pro License, open weights on Hugging Face | This article |
| Tencent Hy4 | 770B parameters | Apache 2.0 | Tencent Hy4 coverage on shattered.io |
| Qwen4 (Alibaba) | Launch discussed with doubts over the open-source promise and a catalog of 470+ open models | Status per our earlier report | Qwen4 launch coverage on shattered.io |
| GLM-5.1 (self-hosted at Revolut) | Reported 10x cut in AI coding costs | Self-hosted deployment | Revolut coverage on shattered.io |
| Gemini 4 Carbon (closed) | 59 on the Artificial Analysis index, $2 input price | Proprietary API | Gemini 4 Carbon coverage on shattered.io |
On raw size, Kimi K3.1 Pro runs about 3.6 times the 770 billion parameter count we reported for Tencent Hy4. Size is not skill, though. Our Qwen4 launch coverage showed how much the open-source label depends on follow-through. Promised weights that do not arrive count for nothing, and Moonshot AI at least has files up on Hugging Face to inspect today.
The closed-model line in the table is there for a reason. Gemini 4 Carbon scored 59 on the Artificial Analysis index at a $2 input price, according to our earlier report. Any open-weight model has to beat that on total cost of ownership, not just on a benchmark, once you count the hardware needed to host it.
Market Impact: Who Gains From a 2.8 Trillion Parameter Open Release #
No price exists for Kimi K3.1 Pro in the sources, so any API-cost comparison would be guesswork. We can still map the likely winners. Inference providers and cloud platforms gain a heavyweight model to host and resell. Enterprises with strict data rules gain a frontier-scale option they can run inside their own walls, if they can afford the iron. Researchers gain a large MoE checkpoint to probe.
Hardware vendors also have reason to pay attention. A 1.56 TB checkpoint in a 4-bit-class format rewards accelerators with large memory pools and native low-precision support. Demand for such parts was already strong, and a flagship open release gives buyers one more reason to specify them.
Self-hosting economics deserve a sober look. The Revolut GLM-5.1 story showed a company cutting AI coding costs by 10x with a self-hosted model, so the idea has proven value at scale. A fintech with an existing GPU fleet and an engineering team is a very different buyer from a three-person startup, and Kimi K3.1 Pro’s size favors the former.
We saw no market-moving price reaction tied to this release in the sources. Open-weight launches rarely move public stocks on their own. The effect shows up more slowly, in cloud catalogs, in procurement talks and in how closed-model vendors set their prices.
Historical Context: How Open Weights Got This Big #
Open-weight releases have climbed steadily in scale. They began as models a hobbyist could run on one graphics card and grew into multi-hundred-billion-parameter systems that need clusters. The shift to Mixture-of-Experts made that growth affordable, since total parameters could balloon while per-token compute stayed in check.
Chinese labs have driven much of that climb. Our October 2026 open-source LLM roundup put Kimi, GLM and DeepSeek side by side, and the Tencent Hy4 and Qwen4 stories sit in the same current. Kimi K3.1 Pro continues that pattern and pushes the total parameter count into the trillions.
The other trend is the move to low-precision distribution. Shipping weights at 4 bits used to mean community quantizations produced after the fact. Here the lab itself ships MXFP4 weights, which signals that training and serving at low precision has become a first-class path and not a workaround.
Five Predictions for Kimi K3.1 Pro #
These are our judgments, not reported facts. We offer them so you can test them against what happens next.
- Hosted versions arrive first. Most developers will meet Kimi K3.1 Pro through an inference provider or aggregator before they ever download a file. The hardware barrier makes that the default path.
- Independent benchmarks decide the narrative. Claims about quality stay thin until outside evaluators publish results. Expect the conversation to shift from parameter counts to measured performance within weeks.
- License details spark debate. A custom license on a flagship open release usually draws scrutiny from legal teams and community lawyers. Clarifications or FAQs from Moonshot AI would not surprise us.
- Smaller derivatives follow. Pressure will build for distilled or pruned versions that fit on modest hardware, whether they come from Moonshot AI or from the community, license permitting.
- Rivals answer with their own open drops. With Qwen4 and Tencent Hy4 already in the conversation, other labs have every reason to counter with releases of their own.
For more model news and comparisons, browse our AI and machine learning section.
What Remains Unconfirmed #
Precision matters more than speed on a story like this, so here is the unconfirmed list in one place.
- Price. No API or hosting price appears in the sources.
- A single official release date. Pages carry dates from November 2 through 5, 2026, and none settles the question.
- On-record quotes. We found no attributable direct statement from Moonshot AI, so this article paraphrases the published model card and press coverage.
- Benchmark results. The sources include no test scores, so we make no quality claims.
- License terms. The license has a name, and we have not verified its contents.
- Supported hardware and inference engines. Neither is described in the material we reviewed.
One more caution. The story headline we worked from refers to Kimi K3.1 Pro model weights. We read that as the weights release for Kimi K3.1 Pro and found nothing to suggest a separate product by that name.
Frequently Asked Questions About Kimi K3.1 Pro #
What is Kimi K3.1 Pro?
Kimi K3.1 Pro is an open-weight, natively multimodal agentic model from Moonshot AI. It uses a Mixture-of-Experts design with 2.8 trillion total parameters and 104 billion active per token. Its identifier is kimi-k3.1-pro.
Where can I download the Kimi K3.1 Pro weights?
The files are hosted under moonshotai/Kimi-K3.1-Pro on Hugging Face. The repository shows a storage size of 1.56 TB, so check your disk space before you start.
How big is the Kimi K3.1 Pro context window?
The listed context length is 1,048,576 tokens, which is 1 million tokens. The sources do not include long-context test results, so real-world quality at that length is unverified.
Does Kimi K3.1 Pro accept images?
Yes. The model card lists text and images as inputs, handled through the MoonViT-V2 vision encoder, which has 401 million parameters. Audio and video input are not mentioned.
Is Kimi K3.1 Pro free for commercial use?
We cannot say. The model ships under the Kimi K3.1 Pro License, and we have not verified its terms. Read the license file in the repository before building a product on the model.
How does Kimi K3.1 Pro relate to Kimi K3?
Moonshot AI post-trained Kimi K3.1 Pro from Kimi K3. The sources do not detail what changed between the two, so test your own workloads before you migrate.
What does it cost to use Kimi K3.1 Pro?
No price appears in the available sources. Self-hosting costs depend on your hardware, and hosted pricing from third-party providers, if any, will be set by those providers.