Kimi K3 Open Weights Are Now Live — Everything Confirmed About the World’s Largest Open-Weight Model

The Kimi K3 open weights are now live on Hugging Face. Moonshot AI has released its 2.8-trillion-parameter model for free download — confirmed today, July 27, 2026 — making it the largest open-weight model publicly available. Developers can download, modify, and self-host the full model weights. A technical report covering architecture, training methodology, and benchmark evaluations is expected to follow shortly.

The release comes eleven days after K3’s API debut on July 16, which sent Kimi K3 to the top of LMArena’s Code Arena and triggered a sharp selloff across Chinese AI competitor stocks.

What the Kimi K3 Open Weights Include

The weights are available at huggingface.co/moonshotai/Kimi-K3 under Moonshot’s verified official organisation. Download only from this source — do not use third-party mirrors. The native release format is MXFP4 safetensors, approximately 594 GB. Community GGUF and BF16 re-quantizations are expected within days.

Alongside the weights, Moonshot has contributed a KDA prefill cache implementation to upstream vLLM, ensuring the model loads reliably across the inference ecosystem. The full technical report is expected to follow the weight release.

The context window is 1 million tokens — meaning users can feed entire large codebases or lengthy documents into a single query. That capability, combined with K3’s documented long-horizon coding strength, is where the model’s practical differentiation lies relative to every other open-weight model available today.

The Architecture: Verified Specs

Kimi K3 is built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), two architectural innovations developed internally by Moonshot. It uses Stable LatentMoE sparsity, activating 16 out of 896 experts per forward pass — meaning the active compute per query is a fraction of the 2.8T headline figure. For context, our pre-launch architecture deep dive covered what the MoE design means for inference efficiency in practice.

Training used quantization-aware MXFP4 weights with MXFP8 activations from the supervised fine-tuning stage onward. Moonshot reports approximately 2.5× improvement in overall scaling efficiency compared to Kimi K2. The model includes native vision capabilities — multimodal from the base, not bolted on.

Where K3 Actually Sits in the Rankings

Moonshot is transparent about this in its official blog. K3’s overall performance still trails Claude Fable 5 and GPT-5.6 Sol — the two most powerful proprietary models currently available. However, it outperformed both companies’ prior-generation models across coding and general agent evaluations. Arena blind evaluations showed developers selecting Kimi K3 ahead of leading US models specifically on front-end coding tasks.

That is the honest benchmark framing — and it matters because most coverage of Chinese frontier models omits the admission entirely.

API Pricing and Availability

Kimi K3 is available via API at platform.kimi.ai using model string kimi-k3. Moonshot’s published pricing:

• Cache-hit input: $0.30 per million tokens
• Cache-miss input: $3.00 per million tokens
• Output: $15.00 per million tokens

The cache-hit rate matters here. Moonshot’s Mooncake disaggregated inference architecture achieves above 90% cache hit rates on coding workloads, making the effective input cost closer to $0.30/MTok for most coding use cases. Beyond the API, K3 is live on Kimi.com, Kimi Work (desktop app v3.1.0+), and Kimi Code via the /model command in terminal.

Hardware Reality for Self-Hosting

You cannot run Kimi K3 on a consumer GPU. The minimum viable self-hosting configuration is approximately 8× H100 80GB GPUs for experimental inference. Moonshot recommends supernode configurations with 64 or more accelerators for production throughput. An RTX 4090, Mac Studio, or any consumer-class machine cannot load the model even with quantization.

Furthermore, none of the three major hyperscale cloud platforms — AWS Bedrock, Microsoft Azure Foundry, or Google Vertex AI — have integrated Kimi K3 into their managed offerings yet, according to Bloomberg Intelligence analysts. Third-party open-model hosting providers are expected to be the primary beneficiaries of today’s weight release in the near term.

Moonshot paused new subscriptions within 48 hours of the July 16 API launch after demand crushed GPU capacity. The open-weight release shifts the compute burden to whoever downloads the weights.

Moonshot’s Business Trajectory

The weight release arrives against a significant commercial backdrop. Since K3 launched on July 16, Moonshot’s daily revenue has grown by a factor of at least six. The company reached $300 million in annual recurring revenue in June — up from $200 million in April — and is seeking a new funding round at a $50 billion valuation ahead of a potential Hong Kong IPO as soon as this year.

K3’s July 16 debut also triggered sharp declines across Chinese AI competitor stocks. Z.ai lost as much as 30% of its value in Hong Kong trading, MiniMax Group fell as much as 16%, and Alibaba dropped 4% — a direct market response to K3’s arena performance on front-end coding tasks.

Washington Scrutiny

The open-weight release arrives under active scrutiny from Washington. White House Office of Science and Technology Policy Director Michael Kratsios accused Moonshot last week of training K3 using banned Nvidia chips and conducting large-scale distillation against US models, including Anthropic’s Fable model. Moonshot has not responded to requests for comment on the allegations.

The accusation is unconfirmed. It is reported here because it is a material fact for any enterprise evaluating K3 for production use — particularly US-based organisations subject to export control compliance.

The Three Limitations Moonshot Published

Moonshot included an explicit limitations section in its official tech blog — uncommon transparency for a frontier model launch. All three are sourced directly from Moonshot.

Sensitivity to thinking history: K3 was trained in preserved thinking history mode. If an agent harness fails to pass back all historical thinking content, or if a session is switched to K3 mid-conversation, generation quality can become highly unstable. Use a verified harness such as Kimi Code and avoid mid-session model switches.

Excessive proactiveness: K3’s training emphasises long-horizon, challenging tasks. When it encounters minor issues or ambiguous user intent, it may make unexpected decisions on the user’s behalf. Applications requiring well-defined behavioral boundaries need explicit constraints in the system prompt or AGENTS.md.

User experience gap: Despite strong benchmark performance, K3 exhibits a noticeable gap in overall user experience compared with Claude Fable 5 and GPT-5.6 Sol. Moonshot states this plainly in the official blog.

Open Weight vs Open Source

Moonshot uses open-source language in its marketing. The accurate term is open-weight. You receive the trained parameters — downloadable, fine-tuneable, commercially deployable under the expected Modified MIT license. Training data and the full training pipeline are not included. For most deployment purposes the distinction is academic, but compliance teams requiring strict OSI definitions should note it.

The China Data Question

Open weights change the self-hosting calculus, but they do not change one fundamental fact: Moonshot AI is a Beijing-based entity subject to China’s National Intelligence Law. For teams using the hosted API, prompts and outputs pass through servers under that legal framework. Self-hosting the open weights on your own infrastructure eliminates that exposure entirely — the primary reason for regulated industries to absorb the significant hardware cost of running K3 locally. Microsoft is already testing Kimi K3 for Copilot, which implies Azure-hosted inference rather than direct Moonshot API routing.

Frequently Asked Questions

Are the Kimi K3 open weights live now?

Yes — confirmed live as of July 27, 2026. Download from the official Moonshot AI Hugging Face repository at huggingface.co/moonshotai/Kimi-K3. Do not use third-party mirrors.

What is the Kimi K3 download size?

Approximately 594 GB for the native MXFP4 safetensors release. Community GGUF quantizations are expected within days — Q4 variants will likely land in the 300–400 GB range.

What hardware do I need to run Kimi K3 locally?

A minimum of 8× H100 80GB GPUs for experimental inference. Moonshot recommends 64+ accelerators in supernode configurations for production throughput. Consumer hardware cannot load the model. Kimi K2.7 Code on Hugging Face is the practical locally-runnable alternative.

What is Kimi K3’s API pricing?

$0.30/MTok for cache-hit input, $3.00/MTok for cache-miss input, and $15.00/MTok for output. The official Kimi API achieves above 90% cache hit rates on coding workloads, making effective input cost much closer to $0.30 for most coding use cases.

Is Kimi K3 better than GPT-5.6 Sol or Claude Fable 5?

No — Moonshot’s own official blog states K3 still trails both models overall. However, K3 outperforms all other tested models and leads on front-end coding tasks in arena blind evaluations. It is the strongest open-weight model at this scale.