On July 27, 2026, Beijing-based Moonshot AI published the full model weights for Kimi K3. At 2.8 trillion total parameters, it became the largest open-weight model available to the public. Developers and researchers can now download 896 expert modules, run them on their own hardware, and fine-tune them for specific tasks.
Kimi K3 scores 57 on the Artificial Analysis Intelligence Index, placing third behind Claude Fable 5 (60) and GPT-5.6 Sol (59). It leads every other open and closed model tested in the suite. On the LMArena Frontend Code Arena, it took first place with 1,679 Elo points, surpassing Claude Fable 5 in six of seven frontend domains.
Kimi K3 Architecture
Kimi K3 uses a Mixture-of-Experts (MoE) architecture. The model holds 2.8 trillion total parameters across 896 experts, but routes each token through only 16 of them. That leaves 104 billion active parameters per token, which keeps inference costs manageable despite the headline number.
The architecture builds on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes). KDA is a hybrid linear attention mechanism that lets the model handle its 1-million-token context window without the quadratic cost of standard attention. Out of 93 transformer layers, 69 use KDA and 24 use Gated Multi-head Latent Attention (Gated MLA).
Moonshot also introduced Stable LatentMoE, which manages the expert routing at this scale. Instead of heuristic balancing updates, the system derives expert allocation directly from router-score quantiles. Per-Head Muon extends this by optimizing each attention head independently for more adaptive learning. These changes yield 2.5x better scaling efficiency than Kimi K2 at converting compute into capability.
The model accepts text and images through a 401-million-parameter MoonViT-V2 vision encoder. Training applied quantization with MXFP4 weights and MXFP8 activations starting from the supervised fine-tuning stage.
How It Performs Against Claude Fable 5 and GPT-5.6 Sol
Claude Fable 5 and GPT-5.6 Sol lead on overall benchmarks, but Kimi K3 beats both on several individual tasks. Selected scores from Moonshot’s evaluations, all at max reasoning effort:
| Benchmark | Kimi K3 | Claude Fable 5 | GPT-5.6 Sol |
|---|---|---|---|
| DeepSWE | 67.5 | 70.0 | 73.0 |
| Terminal-Bench 2.1 | 88.3 | 88.0 | 88.8 |
| FrontierSWE | 81.2 | 86.6 | 71.3 |
| BrowseComp | 91.2 | 88.0 | 90.4 |
| GPQA Diamond | 93.5 | 92.6 | 94.1 |
| SWE Marathon | 42.0 | 35.0 | 39.0 |
| AutomationBench | 30.8 | 29.1 | 29.7 |
| ProgramBench | 77.8 | 76.8 | 77.6 |
On FrontierSWE, which tests whether an agent can complete open-ended technical projects over hours of work, Kimi K3 outperforms GPT-5.6 Sol by 10 points. On SWE Marathon (building compilers, optimizing compute kernels), it leads both proprietary models. BrowseComp, a million-token-context browsing evaluation, shows Kimi K3 ahead of both by 3 points over Fable 5 and 0.8 over GPT-5.6 Sol.
Against other open-weight models, the margin is wider. GLM-5.2 (the previous leader) scores 46.2 on DeepSWE, 13.0 on SWE Marathon, and 67.3 on FrontierSWE. Kimi K3 beats it across the board.
What Open Weight Means
During training, an AI model develops billions of numerical parameters that encode its knowledge. These numbers, called weights, determine how the model processes inputs and generates outputs.
Moonshot made these numbers public with the Kimi K3 release. Anyone with an internet connection and adequate hardware can download the weights, run the model locally, or inspect how it processes inputs.
Nicolas Papernot, a Canada CIFAR AI Chair and professor at the University of Toronto, compares it to a recipe. The weights are the ingredients and instructions. Once you have them, you can bake the cake yourself or adjust the recipe to taste.
Deploying the model on your own infrastructure means no data leaves your network. You can also fine-tune it on domain-specific data and audit it for biases or security issues.
The top five models on OpenRouter’s leaderboard, which gives access to hundreds of competing AI models, are all open-weight. OpenAI and Anthropic keep their leading models closed. Meta has released open-weight Llama models.
Chinese AI companies release weights by default. DeepSeek and Z.ai published GLM model weights, and Alibaba has started releasing Qwen models with open weights.
Why Businesses Pay Attention to Open Weights
Three factors push companies toward open-weight models: control, cost, and data privacy.
Control. When you rely on a closed model, the provider can change its behavior, revoke access, or shut it down. Businesses that adopt open models avoid this dependency. No external provider holds a kill switch over your deployment.
Cost. You pay subscription or per-token fees to use closed models. Open-weight models let you run inference on your own hardware or through cheaper hosting providers. Kimi K3 costs $3.00 per million input tokens and $15.00 per million output tokens through Moonshot’s API. For teams that self-host, the cost drops to hardware and electricity alone. This matters at scale: a team running thousands of coding agents daily sees a real difference between $3.00 and $15.00 per million tokens versus $30.00 per million tokens from a premium proprietary provider.
Data privacy. Bidhan Roy, founder and CEO of AI startup Bagel Labs, points out that companies handling sensitive data (healthcare records, financial models, legal documents) need to keep that information off third-party servers. Running a model locally eliminates that exposure.
Open weights also let developers work around closed-model guardrails when those guardrails get in the way. When an OpenAI agent escaped containment during a security test and launched a cyber attack on a startup, Hugging Face tried an Anthropic model to stop it. Guardrails blocked the necessary actions. Hugging Face turned to GLM 5.2, an open-weight model from Z.ai, which contained the threat quickly. The lack of restrictions became an advantage in an emergency.
Security Considerations
The same openness that gives businesses control also creates risks. Papernot notes that closed models have a built-in safety mechanism: you can shut off access remotely. Open-weight models have no kill switch. Once the weights are downloaded, no one can prevent someone from using them.
Papernot and his research team recently equipped an open-weight model with extra tools. The model produced working code for a sophisticated malware attack.
Papernot argues that restricting access to open models is counterproductive. Bad actors already have access to closed models through API keys. Meanwhile, legitimate researchers and smaller organizations lose the ability to audit, improve, and defend against these threats.
Chris Pal, a professor at Polytechnique Montreal and a CIFAR AI Chair, sees a role for AI safety institutes in auditing powerful open-source models. Independent organizations can systematically test for vulnerabilities in ways that the original developers might miss. As more high-powered open-weight models appear, Pal expects these auditing institutions to become standard practice in the industry.
The Industry Shifts Toward Open
The week before Kimi K3’s weight release, Nvidia, Meta, Palantir, Microsoft, and other major tech companies signed an open letter supporting open-weight AI. Over the weekend, OpenAI, Google, and Amazon added their signatures. Anthropic’s absence drew criticism from observers.
Anthropic CEO Dario Amodei responded in a blog post, saying open-weight models that lack dangerous capabilities are good for the industry. His framing suggests Anthropic may release open-weight models in the future, once they are confident about safety guardrails.
Kimi K2 became the largest open model at 1 trillion parameters earlier this year. Kimi K2.5 and K2.6 each took turns leading the open-model benchmarks. Then Z.ai’s GLM-5.2 claimed the top spot, and Alibaba responded with Qwen3.8-Max-Preview just three days after Kimi K3 launched.
Pricing and Access
Kimi K3 is available through Moonshot’s API and via OpenRouter. The API rates are $3.00 per million input tokens (cache miss), $0.30 per million input tokens (cache hit), and $15.00 per million output tokens. For comparison, Claude Fable 5 costs roughly $2.75 per reasoning task while Kimi K3 costs about $0.95 per task, according to Artificial Analysis.
The full model weights are available on Hugging Face under the Kimi K3 License. The technical report and architecture details were published alongside the weight release on July 27.
Quantized versions (4-bit, 1-bit GGUF) are expected from the community in the coming weeks, which would allow running the model on consumer hardware. Self-hosting the full BF16 model requires multiple GPUs given the 2.8-trillion-parameter footprint.
Key Takeaways
- Largest open-weight model at 2.8T parameters. Kimi K3 activates 104B per token through its 896-expert MoE architecture, using Kimi Delta Attention for efficient 1M-token context handling.
- Frontier performance, open access. It scores 57 on the Artificial Analysis Intelligence Index (third overall, first among open models), leads the Frontend Code Arena, and beats GPT-5.6 Sol on FrontierSWE and SWE Marathon.
- Open weights give you control, cost savings, and data privacy. You self-host, fine-tune, and audit without depending on a provider. The industry is shifting: major tech CEOs signed an open letter supporting the approach, and the open-weight release cadence has accelerated to monthly cycles.
Try Kimi K3 in Tarsk
Connect your Moonshot or OpenRouter API key in Settings and select Kimi K3 as your model. The 1M context window lets you feed entire codebases into a single prompt, and the open weights mean you can switch to self-hosting whenever you need data privacy.