Moonshot releases Kimi K3.1 Pro weights: 1.56 TB under a custom licence

Moonshot AI has published the full weights of Kimi K3.1 Pro on Hugging Face. The release, dated November 4, 2026, is a 1.56 TB download split across 96 safetensors shards. When Unicorn API covered the model’s launch on October 23, the weights were still a promise. Now there is a concrete answer to three questions that were open then: how large the checkpoint is, what hardware can serve it, and what the licence allows.

The short answers are these. The checkpoint is far smaller than many builders expected, because it ships only in the native 4-bit format Moonshot trained it to serve in. AMD has shown it running on one eight-GPU node on day 0. The licence is a custom one, not OSI-approved, and it puts a $20 million revenue line in front of some commercial hosts.

What changed since the October launch

At launch, Kimi K3.1 Pro was available through Moonshot’s API, with open weights promised for later. The repository is now live as moonshotai/Kimi-K3.1-Pro. It is ungated, meaning no access request is needed, and it holds 111 files: 96 weight shards plus the config, the tokenizer, four Python files for the custom architecture, the model card and the licence.

Moonshot describes Kimi K3.1 Pro as “an open-weight, natively multimodal agentic model” and its most capable 3T-class model to date. It is the first model in the K3.1 series and is post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of downstream domains. Reinforcement learning (RL) here means training the model on tasks where it is rewarded for successful outcomes. Moonshot’s API documentation says K3.1 Pro uses “substantially larger post-training compute” than K3.

Why the download is 1.56 TB, not 5.6 TB

For the eleven days between the API launch and the weight release, much of the coverage cited 5.6 terabytes of BF16 weights, according to an analysis by OrcaRouter. The published repository totals 1,561 GB. It is the same model. The difference is precision.

The model card lists the quantization as “MXFP4 weights / MXFP8 activations (quantization-aware training)”. MXFP4 is a 4-bit microscaling number format. Quantization-aware training means the model was trained to work in that low precision, rather than compressed after training. OrcaRouter notes that this is the native serving precision, that the repository carries the compressed-tensors tag, and that no BF16 checkpoint is published alongside it.

AMD’s technical write-up gives a more detailed breakdown. The routed experts are stored in MXFP4, with one E8M0 scale per 32 values. Attention, the indexer, the shared experts, the dense MLP, the latent projections, the router, the LM head and the vision encoder are kept in BF16. Teams sizing storage and memory should therefore plan around the 1.56 TB figure.

Spec	Kimi K3.1 Pro
Architecture	Mixture-of-Experts, 2.8T total / 104B activated parameters
Layers	93 (1 dense); 69 KDA + 24 Gated DSA attention layers
Experts	896, 16 selected per token, 2 shared
Context length	1,048,576 tokens
Modality	Text and image
Vision encoder	MoonViT-V2, 401M parameters
Vocabulary	160K
Quantization	MXFP4 weights / MXFP8 activations, quantization-aware training
Repository	96 shards, 1,561 GB total, ungated

What Kimi K3.1 Pro changes over Kimi K3

The architecture is almost the same as Kimi K3. The model card names one change: “Compared with Kimi K3, we replace the full-attention Gated MLA layers with Gated DeepSeek Sparse Attention (Gated DSA) layers for more efficient long-context modeling”.

AMD explains how this works in practice. Of the 93 decoder layers, 69 use Kimi Delta Attention (KDA), a gated linear attention with a fixed-size state per request. The other 24 are latent-attention layers. In K3.1 Pro, each of those 24 layers gains an indexer that scores the request’s history and selects the top 2,048 tokens. Attention then reads at most 2,048 tokens per query however long the context grows, and only the lightweight indexer scans the whole history. The goal is to make very long prompts cheaper to serve.

OrcaRouter’s reading is that the attention change is there for long-context efficiency, not for a jump in capability, and that the capability gains come from post-training. AMD also describes a separate DSpark speculative-decoding draft: a five-layer model that proposes a block of draft tokens per step for the main model to check, which speeds up generation.

Moonshot also offers a Kimi K3.1 Pro Highspeed variant through its API. The company says it runs at 2-3× the speed of K3.1 Pro.

The licence: permissive opening, three conditions

The weights are released under the Kimi K3.1-Pro License, a custom licence that OrcaRouter notes is not Apache 2.0, not MIT and not OSI-approved, despite a permissive opening. That opening grants permission to use, copy, modify, merge, publish, distribute, sublicense and sell the model, to run, deploy and fine-tune it, and to create derivative works. The conditions follow, as summarised by OrcaRouter:

• Section 2, model-as-a-service: a business that gives third parties access to inference or fine-tuning, with group revenue above $20 million over any consecutive 12 months, needs a separate agreement with Moonshot before commercial use. Two kinds of use are excluded from “model as a service”: end-user products where the model is embedded inside a feature or harness, and services that only relay requests to models hosted by someone else.
• Section 3, attribution: commercial products with more than 100 million monthly active users, or more than $20 million in monthly revenue, must display “Kimi K3.1-Pro” prominently in the interface. This is a branding obligation, not a payment.
• Section 4, exemptions: neither condition applies to internal use, defined as use that does not make the software, its outputs or its capabilities available to third parties.

For most builders, the Section 2 carve-out is the key detail. An app that embeds the model inside a product feature is treated differently from a large inference host reselling raw access. Inference providers above the revenue line must negotiate with Moonshot.

Day-0 serving on AMD Instinct GPUs

AMD published a technical article on release day showing Kimi K3.1 Pro serving on eight AMD Instinct MI355X GPUs, using both ATOM, AMD’s inference engine, and vLLM. Both load the released MXFP4 checkpoint and the DSpark draft. ATOM uses tensor parallelism across the eight GPUs of one node and handles text and image input from the checkpoint without conversion.

AMD reports these memory figures: with the DSpark draft loaded, each GPU peaks at about 199 GB of its 288 GB during start-up, and “a single 1M-token request fits on one node”. AMD’s optimisations include re-quantizing the BF16 attention and dense layers to FP8 at start-up, which takes about 2 seconds for 559 linears per rank, and storing the latent KV and index caches in FP8. On the vLLM side, AMD describes memory-efficient sparse attention with decode context parallelism for long-context serving, plus a specialised decode path for low-latency interactive use. Blockchain.News reports that this vLLM decode path cuts token-generation latency by 27%, according to AMD’s internal benchmarks.

On accuracy, AMD ran Moonshot’s Kimi Vendor Verifier at maximum reasoning effort. It reports an OCRBench score of 0.871 against Moonshot’s reference of 0.872, and 0.991 on tool-call JSON schema validation. These are vendor-run checks that serving on AMD hardware matches Moonshot’s reference outputs. They do not independently measure the model’s capability.

How the benchmarks read, and what they leave open

All capability numbers for Kimi K3.1 Pro still come from Moonshot. The model card reports results at maximum reasoning effort, temperature 1.0 and top-p 0.95. In Moonshot’s documentation, K3.1 Pro scores 65.7% on Terminal-Bench 4.0, ahead of GPT-6 Astra (58.2%) and just behind Claude Opus 5.5 (66.4%). On Moonshot’s internal Kimi Code Bench 3.0 Long-horizon test, it scores 61.2%, up from Kimi K3’s 55.95%.

The same tables show clear gaps. In Moonshot’s own comparison:

• HLE-Full: 51.1, against 59.8 for Claude Opus 5.5 and 57.6 for Claude Fable 5.1.
• HLE-Diamond: 39.4, against 65.8 for GPT-6 Astra.
• Program Bench: 31.5, against 65.0 for Claude Opus 5.5.
• ExploitBench: 74.7, against 100 for GPT-6 Astra; ExploitGym: 20.7, against 42.4.

The cyber scores relate to the earlier article’s main point: Moonshot said it deliberately curbed the model’s hacking skills, and the published table shows K3.1 Pro below the leading proprietary models on both exploit benchmarks.

Several caveats from the footnotes matter for comparisons. K3.1 Pro is evaluated with the Claude Code harness on the coding benchmarks, while many competitor scores are taken from their makers’ blogs, technical reports or public leaderboards rather than rerun under the same conditions. Some rows are internal benchmarks, including Kimi Code Bench, Business Operation and Kimi Hallu Bench.

Moonshot also reports that on its internal Kimi Hallu Bench Hard, the claim-level hallucination rate fell from 23.4% for K3 to 8.3%. The company adds that the test is a deliberately adversarial stress test and that “the reported rates do not represent overall error rates in daily use”. Now that the weights are public, outside groups can rerun these evaluations on their own hardware and harnesses. None of the sources here report such independent results yet.

What it means for builders

• Self-hosting is now possible at a known size. One 1.56 TB checkpoint, in native MXFP4, with no BF16 version to choose instead.
• A single-node path exists on AMD. AMD documents eight MI355X GPUs serving the full 1M-token context, with vLLM support as well as AMD’s own ATOM engine. The sources do not describe a comparable day-0 recipe for other hardware.
• Read the licence before building a hosted service. Embedding the model in a product is treated differently from selling inference. The $20 million revenue line and the attribution rule only apply at large scale.
• The hosted option remains. On the Kimi API Platform, K3.1 Pro is a flagship model unlocked after an account top-up, and the account tier sets rate limits.

What to watch next

Three things will show how this release lands. First, independent evaluations run on the public weights, especially on the reasoning and coding benchmarks where Moonshot’s own tables show K3.1 Pro behind Claude Opus 5.5 and GPT-6 Astra. Second, which inference providers take on the model, and whether large hosts sign separate agreements under Section 2 of the licence. Third, whether serving recipes for other hardware appear, following AMD’s day-0 MI355X support. Moonshot calls K3.1 Pro the first model in the K3.1 series, so further K3.1 models are also likely.