Shortly after midnight Beijing time on November 5, 2026, Moonshot AI announced from its official account that the weights of Kimi K3.1 Pro are open for download, and the following morning's tech press coverage kept to a single consistent story. The moment had been thirteen days in the making: when the model launched in the app and on the API on October 23, the company promised the weights would be open before November 4. That promise has now been kept on schedule. The headline puts 2.8 trillion parameters next to 1.56TB on purpose: the capability ceiling and the download reality are each half of this story, and reading only one side distorts it.
Before anything else, a note on where this piece sits. This site's report on Tencent's Hy4 official release covered a different company's launch in the same window and has nothing to do with this one. The VRAM estimation and deployment-path SOP was written on November 4, when the weights were not yet out; under a strict no-speculation rule it ran the resource math without counting on a release; this piece is its sequel, reporting the event and the fine print. A separate piece in this batch, the hands-on weights access review, handles the practical work of downloading, loading, and deploying, and does not overlap with what follows. The structure below runs through the timeline, the model card, two sets of scores, the size of the download, the license, and the access channels, keeping official, media, and third-party figures labeled separately throughout.
The Timeline: A Deadline Honored
Lay the dates out and the shape of the release becomes clear. On October 23, Kimi K3.1 Pro went live in the app and on the API under the model name kimi-k3.1-pro. The part that invited skepticism was the dated promise attached: in this industry, "weights coming soon" routinely means never, and committing to a specific date pins a grading sheet to your own door. At midnight Beijing time on November 5, 2026, the official account declared the weights available, and multiple tech outlets soon ran confirming reports with consistent figures — enough to settle the question. Delivering on a dated weights promise is not the standard pattern of 2026 launches, and it deserves to be recorded as such.
One timestamp detail is worth flagging. The lastModified field on the Hugging Face repository reads November 2, three days before the announcement. That is not a contradiction but a common engineering arrangement: files are uploaded first, and the repository is opened for access later. This piece follows the official November 5 date; knowing the mechanism means a repo predating an announcement will not be misread as a leak or a fabrication.
The Model Card: 2.8T Total, 104B Active, 1M Context
The official model card positions the release as open-weight, natively multimodal, and agentic, calling it the most capable 3T-class model to date and noting that it was post-trained from Kimi K3. The architecture table reads well item by item: total parameters of 2.8T with 104B active under a mixture-of-experts design; a context window of 1048576, which is to say 1M tokens; attention composed of 69 KDA layers and 24 Gated DSA layers — relative to Kimi K3, the full-attention Gated MLA has been replaced with Gated DeepSeek Sparse Attention; and a 401M-parameter vision encoder.
Put the 104B back inside the 2.8T and the activation ratio works out to roughly 3.7 percent, typical for mixture-of-experts designs: each inference lights up only a fraction of the expert network, capacity is built at 2.8T while cost is paid at 104B. That arithmetic is what makes opening the weights of a model this size viable at all.
The attention change deserves its own paragraph. Kimi K3 used full attention with Gated MLA; K3.1 Pro combines 69 KDA layers with 24 Gated DSA layers, where DSA stands for DeepSeek Sparse Attention. Over very long contexts, attention need not be computed in full for every token; sparsification selects the key positions, and long-context inference cost drops accordingly. The 1M window and the sparse attention reinforce each other: without sparsification a 1M-token window is hardly practical, and without the window sparsification has far less to offer. The 401M-parameter vision encoder is a small fraction of the total, but its placement matters — natively multimodal means visual signals enter during training, not through a bolted-on module.
The precision story carries information too: weights are stored in MXFP4 and activations run in MXFP8, which the card describes as quantization-aware training. The phrase means quantization's effects were built in during training rather than applied as an afterthought. Those two lines explain where the 1.56TB comes from and why a quantized model could ship as the release form; 1.56TB is already the post-quantization size, and that fact determines how hard the constraints below really are.
Two Sets of Scores That Must Never Be Mixed
The benchmarks have to be told apart first, because mixing them is the single most common way coverage of a release like this goes wrong.
The first set is the official model card, self-reported: a Terminal-Bench 4.0 score of 65.7, evaluated with the Claude Code harness; the same table lists GPT-6 Astra at 58.2 and Claude Opus 5.5 at 66.4 as comparison points. Alongside sit Terminal-Bench-Science 0.1, SWE-Marathon v1.1, and other internal and public benchmarks. Every time these numbers are quoted, the qualifier "official" belongs in the sentence — a company's own score, on its own harness, with its own choice of comparisons, reads best as a capability sketch rather than an objective ranking.
The harness deserves its mention because the same benchmark can score several points apart under a different evaluation scaffold, and benchmarks grading task execution in realistic working environments are especially sensitive to tool-calling and error-handling design. The official 65.7 holds only under the Claude Code harness as reported in the official card; quoting it without those conditions distorts it.
The second set comes from the third-party Agent Arena, a leaderboard built from real user agent sessions: first among open models, with a net improvement rate of 5.29 percent, ahead of second-placed Xiaomi's MiMo-V2.6-Pro at 2.68 percent, fourteenth on the overall board, from a sample of roughly 17,500 to 18,000 real sessions. Two qualifiers are mandatory here. First, the Arena itself marks the ranking as preliminary, so the phrase "first among open models" is only valid attributed to the Agent Arena and labeled as a third-party preliminary ranking; stripped of that label it becomes an absolute claim the evidence cannot carry. Second, the Arena's net improvement rate and the Terminal-Bench score are two entirely different instruments with different evaluation setups; they must never be placed in the same sentence as comparable quantities, and no cross-leaderboard conversion exists between them.
The Arena's sample comes from real user sessions rather than examiner-written tasks: what users asked, where the model got stuck, and whether it satisfied them all live in the records. A sample of 17,500 to 18,000 sessions is not small, but "preliminary" means the ranking will keep shifting as sessions accumulate. At 5.29 percent against 2.68 percent, K3.1 Pro's net improvement rate is nearly double that of MiMo-V2.6-Pro — that is what "first among open models" means here. Fourteenth on the overall board says a line of closed models still sits ahead; both rankings together give an honest position. These qualifiers are the hardest lines in the reporting brief this piece is built on.
About 1.56TB: What the Number Means in Practice
The size is the figure that travels furthest in coverage: about 1.56TB across 96 weight files, a pair of numbers reported consistently by multiple outlets; a direct capture of the Hugging Face repository shows 112 files in total, the difference being non-weight items such as the README and assorted assets. This piece keeps the "about" on the size figure and credits the file count to media reporting, and readers repeating the numbers are advised to do the same.
Translated into download time: at an idealized full-rate hundred-megabit connection, 1.56TB takes on the order of a day, and a typical home connection stretches it into multiple days, with comparable disk space still needed for checksums and staging. These conversions come from no official source — they are arithmetic on the stated size — but they measure the distance between "downloadable" and "downloaded."
The practical constraint is just as blunt. MXFP4 is already a quantized format, so 1.56TB is the compressed volume; but with 2.8T total parameters and 104B active, the arithmetic in the VRAM estimation SOP makes clear that even quantized, the model at this stage is realistically loaded on cloud GPU clusters, and consumer-grade hardware cannot hold it in full. For most developers the realistic paths reduce to two: rent cloud GPUs and serve the model yourself, or call the API directly. The wording deliberately leaves room — the claim is "at this stage," not "never," because inference engines and hardware keep moving, and any such judgment can be revised by the next generation of both.
The License: Permissive Grants Wrapped in Three Conditions
The license is the Kimi K3.1 Pro License, listed on Hugging Face with the license field set to other. Going by the text of the LICENSE itself, the grant is genuinely permissive: use, modification, distribution, and sale are all allowed. Wrapped around that grant, though, are three conditions, and each one has to be read exactly.
The first is the Model as a Service condition: if the licensee and its affiliates together exceed 20 million US dollars in total revenue over any consecutive 12-month period, commercial use of the model requires a separate agreement with Moonshot AI. Two details matter. The revenue yardstick is a rolling "any consecutive 12 months," not a fiscal year, and the count includes affiliates rather than only the contracting entity. The license also draws an explicit boundary: merely relaying requests to a model hosted by someone else does not constitute MaaS, so a pure pass-through gateway falls outside this clause, while resale-style hosting still needs to be checked against the original text.
The MaaS clause matters because hosted resale — putting the model on a cloud and charging per call — is the most common business built on open weights, and a 20-million-dollar threshold lets smaller players pass while calling larger ones to the negotiating table. What is unusual is how precisely the boundaries are written: the rolling 12-month window, the consolidation of affiliates, and the mere-relaying carve-out all sit in the original text. Run the numbers by team size: a startup earning a few million dollars a year trips none of the conditions; a mid-sized company near the line needs rolling monitoring on the "any consecutive 12 months" yardstick; a product approaching 100 million monthly active users already has the attribution condition waiting.
The second is the attribution condition: a commercial product exceeding 100 million monthly active users or 20 million US dollars in monthly revenue must prominently display "Kimi K3.1-Pro" in its interface. This is the condition most often misread — the same 20-million-dollar figure triggers a separate-agreement requirement under the MaaS clause and a display requirement under the attribution clause, one number with two different consequences, and any citation should keep the two apart.
The third is the carve-out: internal use, or access through Moonshot's own official products and certified inference partners, sits outside the reach of the conditions above.
Channels and Inference Engines: Both Roads Are Open
On the API side, the official entry point is platform.kimi.ai, with the model name kimi-k3.1-pro, compatibility with both the OpenAI and Anthropic protocols, and a reasoning_effort parameter offering low, high, and max tiers. The tiers are worth a sentence: low buys response speed, max buys reasoning depth, high sits between, so one model can be called at different depths by task difficulty. Combined with the 1M window, the design suits long-document analysis particularly well. Readers who want to compare what a token actually costs across providers can turn to this site's API pricing comparison review, which puts same-generation pricing on one table.
On the self-hosting side, the company recommends two inference engines, with official recipes linked for vLLM and an official cookbook for SGLang. Read together with the license carve-out, going through a certified inference partner is among the lowest-friction compliant routes for commercial teams; for anyone who genuinely wants to pull the weights and plan a cluster, the hands-on portion is left to the weights access review in this batch, and this piece will not duplicate it. Nor does one road replace the other: the API fits validation phases and spiky traffic, while self-hosting fits workloads with hard requirements that data never leaves a controlled environment.
To close: what actually landed on November 5, 2026 is a promise kept on schedule. The weights are downloadable, the license is readable, and the paths are documented. Quote the results table with its qualifiers attached — "first among open models" only alongside "third-party preliminary ranking," the 65.7 on Terminal-Bench 4.0 only as an official figure, the "about" kept in front of 1.56TB, and the license conditions cited from the LICENSE text with the MaaS threshold and the attribution condition never merged into one. As for what these weights eventually produce — fine-tuned variants, distilled smaller models, new inference optimizations — the answer will come from the community doing the work, not from the announcement.
Related reading