Moonshot AI published the Kimi K3.1 Pro open weights on Hugging Face on 4 November 2026, and the download is roughly 1.6TB across about 96 files (FourWeekMBA). Our verdict: for most teams that want to run a model themselves, DeepSeek V4.1 Flash 1017 (763B parameters, MIT licence) is the better pick, because Kimi K3.1 Pro needs a multi-node cluster and carries commercial conditions. Kimi K3.1 Pro wins for organisations that already operate large GPU clusters, or that will call it through a hosted API and want the strongest agentic coding results in the open-weight field.
TL;DR
- Moonshot released the weights on 4 November 2026 at moonshotai/Kimi-K3.1-Pro: 2.8T total parameters, 104B active per token, 1M-token context.
- Against Kimi K3, the footprint is the same. The change is the attention stack and additional reinforcement-learning post-training.
- The licence is MIT-style with three extra conditions. It is not OSI-approved open source.
- At about 1.6TB it will not fit on one 8xH200 node, so self-hosting means several machines.
- DeepSeek V4.1 Flash 1017 is the pragmatic self-hosting default; Kimi K3.1 Pro is the better choice if you have the cluster or use a hosted endpoint.
What shipped with the Kimi K3.1 Pro open weights release?
Moonshot describes the model as an open-weight, natively multimodal agentic model, post-trained from Kimi K3 and the first in the K3.1 series (model card). The card lists the following:
- 2.8T total parameters with 104B activated per token
- Mixture-of-Experts with 896 experts, 16 selected per token plus 2 shared experts
- 93 layers, a 160K vocabulary and a context length of 1,048,576 tokens (model card)
- MXFP4 weights and MXFP8 activations, trained with quantization awareness
- MoonViT-V2 vision encoder at 401M parameters, accepting text and images
Because the weights are already quantized, there is no smaller "official" variant to fetch. The repository is the 1.6TB version, and the shards are about 17GB each (example shard). Huawei has also announced inference support through vLLM Ascend and SGLang on Atlas A3 and Ascend 950 systems (TechNode).
What changed compared with Kimi K3?
Very little in size, a fair amount in behaviour. Kimi K3 shipped its weights on 27 July 2026 with the same 2.8T total and 104B active figures (Kimi blog). The model card for K3.1 Pro says full-attention Gated MLA layers were replaced with Gated DeepSeek Sparse Attention layers to make long-context work more efficient. The stack is now 69 KDA layers and 24 Gated DSA layers. Moonshot also reports large-scale joint reinforcement learning across coding, 3D creation, financial research and long-running agent tasks.
The practical reading is that you should not expect a smaller or cheaper model to run. You should expect the same hardware bill with better long-context efficiency and stronger agent performance.
How does Kimi K3.1 Pro perform on benchmarks?
The numbers below are self-reported by Moonshot at maximum reasoning effort (model card). Treat them as vendor claims until independent runs accumulate.
| Benchmark | Kimi K3.1 Pro | Claude Opus 5.5 |
|---|---|---|
| Terminal-Bench 4.0 | 65.7 | 66.4 |
| SWE Marathon v1.1 | 51.7 | 46.3 |
| DeepSWE v1.1 | 75.5 | 74.2 |
| OSWorld V2.1 | 79.6 | 81.8 |
| Program Bench | 31.5 | 65.0 |
The picture is mixed. Kimi leads on the two long-horizon software engineering suites and trails on computer use and Program Bench, where the gap is wide. A third-party tally counts 16 wins out of 37 tests against Opus 5.5 (FourWeekMBA). For a fuller head-to-head, see our comparison of Kimi K3.1 Pro and Claude Opus 5.5.
What are the Kimi K3.1 Pro licence rules?
The licence is an MIT-style grant covering use, copying, modification, distribution, sale, deployment and fine-tuning, with three added conditions (LICENSE file). It is term-identical to the Kimi K3 licence apart from the model name (K3 LICENSE). This is a plain-language summary, not legal advice.
- Model as a Service threshold. If you or your affiliates sell third-party API access to inference or fine-tuning, and aggregate revenue passes 20 million US dollars in any consecutive 12 months, you need a separate agreement with Moonshot before commercial use (LICENSE file).
- Attribution threshold. If a commercial product has more than 100 million monthly active users, or more than 20 million dollars in monthly revenue, the licence requires a prominent attribution (LICENSE file).
- Exemptions. Internal use, and access through Moonshot's official products or certified inference partners, are exempt.
A start-up fine-tuning the model for its own product is unaffected by the first two clauses at current scale. A hosting company reselling an inference API is the case the first clause targets. Hugging Face labels the licence "other" (name kimi-k3.1-pro), which means it is not OSI-approved open source, even though the weights are open.
How do you deploy Kimi K3.1 Pro?
The model repository documents several routes (Hugging Face):
# vLLM: OpenAI-compatible API on localhost:8000
pip install vllm
vllm serve "moonshotai/Kimi-K3.1-Pro"
# SGLang
pip install sglang
python3 -m sglang.launch_server --model-path "moonshotai/Kimi-K3.1-Pro" --host 0.0.0.0 --port 30000
# Docker Model Runner
docker model run hf.co/moonshotai/Kimi-K3.1-Pro
The transformers route needs trust_remote_code=True, the model class kimi_k31 and an upgraded transformers and accelerate. Reasoning depth is set with a top-level reasoning_effort request field, with values low, high or max. The default is max, so set it explicitly if latency matters.
Hardware is the real obstacle. An 8xH200 node offers 1.1TB of HBM3e, which is less than the 1.6TB of weights alone. As a precedent, a community W4AFP8 build of Kimi K3 ran on 16 H200 GPUs across two nodes using SGLang with TP16 and EP16 (vessl/Kimi-K3-W4AFP8, via K3 page; Morph). Expect a similar multi-node setup here, plus room for the KV cache.
If you do not want to run that yourself, Together AI serves the model through Hugging Face Inference Providers, and Moonshot offers an OpenAI- and Anthropic-compatible API at platform.kimi.ai under the name kimi-k3-1-pro.
What should you run if 1.6TB is out of reach?
DeepSeek V4.1 Flash 1017 is the sensible fallback. It has 763B parameters, which is under a third of K3.1 Pro's 2.8T, and an MIT licence with no revenue or attribution clauses (model page). ArtificialWatch ranks it the highest open-weights model on its Bench Index, with a score of 68 and 11th place of 44 overall, and GLM-5.3 (753B, custom licence) sits third (ArtificialWatch).
So the verdict holds. If your constraint is a single server, a simple licence and predictable operations, take DeepSeek V4.1 Flash. If your constraint is the best open agentic coding quality and you have a cluster or a hosted endpoint, take Kimi K3.1 Pro. For wider context, see Qwen vs Claude vs Kimi, the pricing comparison in Ripple 65B against open-weight coding models, and our look at open-source agent platforms for coding.
The release cadence is also worth noting. Tencent launched Hy4 through its API on the same day and says it plans to open-source the weights without giving a date (FourWeekMBA). Our Tencent Hy4 vs Mistral Large 4 piece covers that side.
FAQ
Q: Is Kimi K3.1 Pro really open source?
A: The weights are public and the licence is permissive, but it is not OSI-approved. It adds conditions on large Model as a Service businesses and on very large commercial products, so "open weights" is the accurate term (LICENSE file).
Q: How big is the Kimi K3.1 Pro download?
A: About 1.56 to 1.6TB across roughly 96 safetensors shards of around 17GB each, according to a count of the repository files.
Q: Can Kimi K3.1 Pro run on a single 8xH200 server?
A: Not at the published precision. The weights alone, about 1.6TB by one file count, exceed the 1.1TB of HBM3e on such a node, and a comparable quantized Kimi K3 build needed two nodes of 16 H200 GPUs (community build).
Q: Can I use Kimi K3.1 Pro commercially?
A: Yes, under the licence (Moonshot LICENSE). A separate agreement is needed only if you resell inference access and pass the 12-month revenue threshold in clause 2, and attribution applies only above the clause-3 user and revenue thresholds.
Q: What is the difference between Kimi K3 and K3.1 Pro?
A: The size is unchanged. K3.1 Pro swaps 24 of 93 attention layers to Gated DeepSeek Sparse Attention and adds large-scale reinforcement-learning post-training on coding and agent tasks.
Q: Which model should I self-host instead?
A: For most single-node setups, DeepSeek V4.1 Flash 1017 under the MIT licence is the practical choice.
Corrections and AI disclosure
This article was researched and drafted with AI assistance, then fact-checked against the primary sources linked inline. Corrections are logged and updated as they arrive. See how we work.
Top comments (0)