News · 2026-11-05

Kimi K3.1 Pro’s 1.56 TB release gives providers a new model to host

Moonshot has released Kimi K3.1 Pro’s downloadable checkpoint, making a flagship model available for organizations to inspect, adapt and host on their own infrastructure. The published repository is 1.56 TB, and documented serving runs establish that the weights work on substantial hardware without independently proving every capability claim made before release.

Key facts

• The public weights-release signal is November 4; the repository’s November 2 creation is a different event.
• The full repository is listed at 1.56 TB across 96 weight shards.
• AMD reports serving it on eight 288 GB MI355X accelerators, with about 199 GB occupied per card at startup including a draft model.
• The primary artifact is Moonshot’s Kimi K3.1 Pro model repository.

Ground Truth’s earlier launch report recorded the November weights promise while the checkpoint was still pending. The release changes who can supply the model. A hosted service puts the operator between users and the computation. Downloadable weights let another provider run the system, or let a large organization keep inference on its own infrastructure. That can change bargaining power and customization options even when the hardware is far beyond a desktop setup.

Moonshot’s technical announcement explicitly discusses an “open-weight release.” The wording matters. The files are available, but the custom license carries conditions. It permits broad use, modification and commercial activity, while requiring a separate agreement for qualifying model-service businesses above $20 million in aggregate revenue over a consecutive twelve-month period. Internal use and official or certified-partner access have stated exceptions. A separate large-service attribution condition also applies.

The architecture helps explain the size. K3.1 Pro is a mixture-of-experts model: for each token, a router activates a subset of its many specialist components. Its active parameter count is much smaller than its total parameter count. Think of a large hospital where a patient visits only the relevant departments. The hospital still has to maintain the other departments for different patients. Sparse computation reduces work per token without making the stored specialists disappear.

Disk storage and runtime memory are separate costs. The file listing establishes the repository’s storage footprint. AMD’s release-day account establishes one measured accelerator configuration. It is not a universal memory minimum, and the startup figure includes another model used to accelerate generation. AMD also reports that a million-token request fit on that node.

That one-node deployment corrects an easy misreading of an experimental eight-server Ascend recipe. Different runtimes, hardware and serving designs produce different configurations. Neither recipe justifies saying that eight servers are always necessary, or that the model is practical on ordinary consumer hardware. The dossier contains no independently demonstrated, useful-speed desktop run of this release.

There is also a less visible efficiency change. vLLM’s implementation account describes an indexer that selects relevant historical positions before full attention is computed over them. This resembles consulting a library catalog before opening selected pages. It reduces long-context memory traffic without removing the storage burden of the model’s experts. The underlying mixture-of-experts concept and context memory explain different parts of that bill.

Capability evidence remains more conditional than release headlines suggest. Artificial Analysis’s October evaluation projected a much smaller gap between open and proprietary models if the eventual weights matched the model it evaluated. That was before this checkpoint became public. Its current index also uses a changed evaluation suite, so comparing old and new index numbers does not establish a regression.

AMD and vLLM do report post-release execution and selected correctness checks. Both worked closely with Moonshot. Their evidence addresses whether the checkpoint serves correctly and whether runtime changes preserve selected behavior; it does not supply an independent, broad comparison with rival flagship models. Moonshot’s own benchmark table shows strengths and weaknesses rather than universal leadership.

The release also received a deliberate training intervention. Moonshot says cyber-related material was removed from later training stages for the open checkpoint. The source does not map every hosted endpoint to a specific filtered or unfiltered version. It therefore reinforces the need to test the actual release revision rather than assume it is identical to an earlier service.

The strongest skeptical reception in the release discussion concerns license freedoms, hardware and benchmark comparability. A linked contamination allegation could not be checked and is not evidence here. The durable news is narrower and significant: infrastructure owners now have another capable model to operate themselves, while independent quality and total-cost comparisons remain outstanding.

Primary source, verified: read the paper →

Ground Truth. (2026, November 5). Kimi K3.1 Pro’s 1.56 TB release gives providers a new model to host. Ground Truth. https://groundtruth.day/news/kimi-k3-1-pro-weights-release-hardware-and-evidence.html

@misc{groundtruth:kimi-k3-1-pro-weights-release-hardware-and-evidence,
  title  = {Kimi K3.1 Pro’s 1.56 TB release gives providers a new model to host},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {nov},
  url    = {https://groundtruth.day/news/kimi-k3-1-pro-weights-release-hardware-and-evidence.html}
}

Topics: open-weights · models · inference · moonshot · infrastructure