On November 4, 2026, Moonshot AI released the trained weights for its AI model Kimi K3.1 Pro. The model had been available through apps and an API since October, and users can now download it and run it in their own environments. GIGAZINE's coverage of the post on Moonshot AI's official X account also confirms the November 4 release, and the distribution files can be downloaded from Hugging Face.
Kimi K3.1 Pro uses a Mixture of Experts (MoE) architecture, in which only some of the parameters are used to compute each token. This structure makes it possible to run a very large model while keeping computation down, but the number of parameters actually computed is different from the memory needed to hold the entire model. Anyone considering an in-house deployment needs to check not only the model's scale but also inference engine support, throughput, and license terms.
About 3.7% of 2.8 trillion parameters are used per token, but the download is 1.56TB
Kimi K3.1 Pro has 2.8 trillion total parameters, and 104 billion parameters are used to process each token. Moonshot AI's official model card lists the two figures separately. Total parameters indicate the scale of the whole model, while active parameters indicate how much is actually computed for each token. MoE can reduce the required computation, but the memory burden of holding the whole model does not shrink in proportion.
Dividing 104 billion by 2.8 trillion gives an active-parameter share of about 3.7% per token. However, the total repository size shown on the official Files page on Hugging Face comes to about 1.56TB. This is a rounded figure for the full set of distributed files as of November 5, and it does not indicate the GPU memory (VRAM) required at runtime.
The model uses low-precision numeric formats: MXFP4 for weights and MXFP8 for activations. According to the model card, it uses Quantization-Aware Training, which accounts for the effects of quantization during training, so the weights were not compressed by third parties after release. The distribution includes weight files split into 96 parts, such as "model-00001-of-000096.safetensors." Even in this low-precision form, the size reaches the terabyte range, so required VRAM, the minimum number of GPUs, and operating costs cannot be calculated from the active-parameter share alone.
A new sparse attention mechanism aimed at more efficient long-context processing
Kimi K3.1 Pro also changes the attention mechanism used to reference information in long inputs. Moonshot AI says it replaced the full-attention "Gated MLA" used in the previous-generation K3 with "Gated DeepSeek Sparse Attention (Gated DSA)," which narrows down the information referenced, in order to make long-context processing more efficient. Of the 93 layers, 69 are Kimi Delta Attention (KDA) and 24 are Gated DSA. As a result, K3 and K3.1 Pro require different attention configurations from inference engines.
The maximum context length is 1,048,576 tokens, and the model accepts text and image inputs. Gated DSA is configured with Top-K=2048 to select what to reference. However, the maximum context length and attention settings do not guarantee actual speed on long inputs. The official model card also does not provide measurements that isolate the speedup or performance contribution of the attention change alone.
Moonshot AI recommends three inference engines: vLLM, SGLang, and TokenSpeed. Having the weights does not mean existing environments built for K3 can be used as is; support for the new attention mechanism and numeric formats is required. Before deploying, users need to confirm that their inference engine supports the model's architecture and then measure throughput for their expected input lengths and number of concurrent users. Being able to accept roughly one million tokens of input is a separate matter from processing that input with practical latency.
The model's capabilities also vary by evaluation. In Moonshot AI's official announcement, it scored 65.7 on Terminal-Bench 4.0, but only 31.5 on Program Bench, below the 50.0 for GPT-6 Astra and 65.0 for Claude Opus 5.5 in the same comparison table. Kimi K3.1 Pro's evaluations generally used reasoning_effort=max, temperature=1.0, and top-p=0.95. These are results published by Moonshot AI, and not all of them have been independently reproduced by third parties. Benchmark scores also cannot be used to judge actual inference speed or concurrency.
Huawei Ascend support is in place, but throughput verification is still to come
Chinese outlet IT Home reported, citing an announcement from Huawei's official computing account, that inference for Kimi K3.1 Pro was supported on vLLM Ascend and SGLang from the day of release. The supported products named are the Atlas 800 A3, Atlas 900 A3 SuperPod, and the Ascend 950 supernode series. Being able to deploy the model in these environments matters for those considering options other than NVIDIA GPUs, but equivalent performance has not been confirmed on every model.
The specific verification status can be seen in SGLang GitHub issue #48348, created by iridiumine on November 4. The configuration described is 3 nodes with 8 cards each, for 24 cards in total, with parallelism set to TP8×PP3 (tensor parallelism 8, pipeline parallelism 3). However, 24 cards is the configuration used in this verification, not a minimum requirement for running Kimi K3.1 Pro.
The issue reports that, in this configuration, accuracy has been verified for input processing (prefill) and generation (decode) of a single request. Throughput, meaning the number of requests or tokens processed in a given time, is to be covered by benchmark results added later. In other words, normal inference has been confirmed on a specific Ascend setup, but capacity under many simultaneous users is not yet known.
In real operation, the required performance differs between a small number of users processing long inputs and many users sending short requests at once. A list of supported products and single-request verification results are therefore not enough to estimate speed or cost at deployment. Assessing the practicality of Ascend environments requires measurements matched to expected input lengths and concurrency, and the current results do not allow a judgment on performance or cost relative to NVIDIA environments.
Weights are free to download, but commercial use carries conditions such as revenue thresholds
The Kimi K3.1 Pro weights can be downloaded free of charge, and use, modification, redistribution, and fine-tuning are permitted. However, they are covered by Moonshot AI's custom license, and not every commercial use is allowed unconditionally. Users must retain copyright and license notices and comply with applicable laws, and additional obligations arise depending on the type of business and its revenue.
Article 2 of the license sets special conditions for those operating a "Model as a Service" business. If the user or its affiliates operate such a business and their combined revenue exceeds US$20 million over 12 consecutive months, a separate agreement with Moonshot AI is required before commercial use. The revenue threshold is based on the combined revenue of the user and its affiliates, not on the model-provision business alone.
Model as a Service includes businesses that provide inference or fine-tuning through an API or similar means in a way that lets third parties substantially control inputs, parameters, or training data. On the other hand, end products that embed the model in specific features, and services that merely relay requests to a model hosted by another company, are excluded from this definition. The key question is therefore not simply whether an API is used, but what functionality is offered to third parties and who controls the inputs and settings.
Article 3 also sets a display obligation for large commercial products and services. If a product or service has more than 100 million monthly active users or monthly revenue above US$20 million, "Kimi K3.1-Pro" must be displayed prominently in the user interface. This differs in both business scope and measurement period from the 12-month combined revenue in Article 2, so the two should not be confused.
Article 4 provides exceptions to these conditions. Articles 2 and 3 do not apply to internal use that does not provide the model itself, its outputs, or underlying capabilities to third parties, nor to use through Moonshot AI's official products or authorized inference partners. Even for internal systems, if third parties are given access to the model's outputs or functions, it is necessary to check carefully whether the use falls within the definition of internal use.
With the Kimi K3.1 Pro weights released, users can now choose an operating environment tailored to their own requirements in addition to the official API. However, neither the model's 2.8-trillion-parameter scale nor the roughly 3.7% activation ratio is enough to determine the equipment or costs needed for deployment. As support and measured performance for each inference engine become clearer, comparisons of operating environments are likely to advance. Verifying throughput for your own use case and checking the license terms that apply to how you offer the model are prerequisites for any deployment decision.