Moonshot AI has released the weights for Kimi K3.1 Pro in open access. This is the first model in the K3.1 series, created based on Kimi K3 using large-scale joint RL across hundreds of tasks. It is a 2.8 trillion parameter MoE model, with approximately 104B active parameters per token, a context length of up to 1,048,576 tokens, and a hybrid Gated DSA attention mechanism. The weights are published on Hugging Face along with ready-made deployment recipes for vLLM and SGLang, and the model is also available via API on platform.kimi.ai.

image
image

!image%20active%20%C2%B7%20MOE%20%C2%B7%201024K%20ctx&version=vLLM%200.34.0%2B&path=%2Fmoonshotai%2FKimi-K3.1-Pro)

What happened

Moonshot AI published the weights for Kimi K3.1 Pro, the first release of the new K3.1 series. The model belongs to the MoE class: 2.8 trillion total parameters, approximately 104B active per token, 896 experts (16 active per token), and MXFP4 quantization for expert weights; the context size is 1,048,576 tokens. The main architectural change affected the attention mechanism: in 24 of 93 layers, full Gated MLA attention was replaced with Gated DSA (DeepSeek Sparse Attention) featuring a top-2048 indexer, which accelerates long-context processing. The vendor reports the following results: 65.7 on Terminal-Bench 4.0, 94.2 on GPQA-Diamond, 92.1 on BrowseComp, and 79.6 on OSWorld V2.1. The weights are hosted in the moonshotai/Kimi-K3.1-Pro repository on Hugging Face. Deployment recipes for vLLM and SGLang with TP8, DCP8, and NVFP4 KV configurations have been published, and the model is available via API on platform.kimi.ai under the identifier kimi-k3.1-pro.

Context

The K3.1 series is built as post-training of the existing Kimi K3 base using large-scale joint RL across hundreds of tasks. Therefore, the primary capability gains are expected in agentic scenarios where the model operates in a loop with tools. The Gated DSA mechanism is named after DeepSeek Sparse Attention and references a previously published technique by DeepSeek. The novelty of this release lies not in inventing the method, but in transferring and integrating sparse attention into open frontier weights. This is done in a hybrid manner—far from all layers are replaced—which appears to be a conservative way to achieve efficiency without abandoning full attention entirely. K3.1 Pro continues the notable trend of open frontier models: extreme sparsity (896 experts with 16 active per token) combined with native MXFP4 quantization of expert weights, which makes a million-token context economically viable. The profile of the reported benchmarks—terminal, browser, computer environment, and scientific questions—deliberately covers the strengths of RL post-training, although the near-saturated GPQA-Diamond benchmark is losing its discriminative power.

Why this matters for the industry

Replacing full Gated MLA attention with a sparse mechanism in open frontier weights is a concrete step toward making DSA a standard technique for long contexts: the cheaper the KV-cache, the more practical a million-token context becomes. Ready-made vLLM and SGLang recipes with TP8, DCP8, and NVFP4 KV configurations lower the barrier to entry for self-hosting: open frontier models rarely come with deployment instructions out of the box. For product teams, these weights turn 1M-context into a working option: agentic scenarios involving terminals, browsers, and computer environments, as well as working with long documents and repositories, no longer require pre-fragmenting data. In the economics of agentic products, this is a significant shift: long context is no longer a premium feature exclusive to closed vendors, and the protective moat for startups shifts from the thesis of "we have the best model" to orchestration, vertical data, and distribution. The key choice for builders is formulated as "API or self-host." The availability of both the base Kimi K3 and the post-trained K3.1 Pro in open weights additionally provides the community with material to study the contribution of RL post-training.

Why this matters for users

The weights can be downloaded from Hugging Face and run independently, but the barrier is high: a node with 8 or more B300 or GB300-level GPUs is required, so for most teams, local deployment remains a future option. The fastest way to test is the API on platform.kimi.ai with the identifier kimi-k3.1-pro: without hardware costs, you can run your agentic and coding scenarios (terminal, browser, long documents) against your current stack and compare quality, speed, and price. For readers, this is one of the strongest open models for agentic tasks and code available for local deployment, and the million-token context means that entire code archives and long conversations can be fed to the model in full, without manual chunking.

What is currently unknown / limitations

Benchmark figures are currently vendor-reported: the provided sources do not include independent reproductions, leaderboard measurements, or latency and throughput measurements under real loads. The breakdown of RL post-training into specific tasks is not provided in the sources, so the contribution of this specific methodology cannot be assessed. API prices are absent from the sources, so economic comparisons with self-hosting must be verified independently. Hardware requirements are strict: local deployment is unavailable without a cluster of 8 or more B300 or GB300-level GPUs. Finally, assessments that hybrid DSA will become the standard for long contexts in open models are extrapolations, dependent on the confirmation of vendor results and the maturation of DSA and NVFP4 KV support in vLLM and SGLang.

Sources

Author

Look at AI, editorial team