The open-weight AI race just got a new leader.
On July 27, 2026, Moonshot AI released the full weights of Kimi K3 on HuggingFace. With 2.8 trillion parameters, it is the largest open-weight model ever released. It comes with a 1-million-token context window, native vision support, and a new architecture called Kimi Delta Attention.
Here is the full breakdown of what this model is, how it performs, and what it means for developers.
What Is Kimi K3?
Kimi K3 is an open-weight, native multimodal model built by Moonshot AI, a Chinese AI lab based in Beijing. They describe it as their "most capable model to date" and the "world's first open 3T-class model."
The model was announced on July 16, 2026 with a promise to open-source the weights by July 27. They delivered on that promise right on schedule.
Key specs:
- Architecture: Mixture-of-Experts (MoE) with 896 experts, 16 activated per token
- Total parameters: 2.8 trillion (2.8T)
- Activated parameters: 104 billion per forward pass
- Context window: 1,048,576 tokens (1 million)
- Modality: Text + Image (native, not stitched)
- Vision encoder: MoonViT-V2 (401M parameters)
- Quantization: Native MXFP4 weights / MXFP8 activations (trained this way)
- Attention mechanism: 69 layers of Kimi Delta Attention (KDA) + 24 layers of Gated MLA
The model uses a Stable LatentMoE framework that activates 16 out of 896 experts per token. Moonshot claims this gives roughly 2.5x improvement in scaling efficiency over their previous Kimi K2 model.
How Does It Perform Against Claude and GPT?
This is where it gets interesting. Kimi K3 goes head-to-head with the frontier models from Anthropic and OpenAI.
On the HuggingFace model card and tech report, Moonshot published benchmark comparisons against Claude Opus 4.8, Claude Fable 5, GPT-5.5 (xhigh), GPT-5.6 Sol, and GLM-5.2.
Here is how it stacks up on key benchmarks:
Reasoning and Knowledge:
- GPQA Diamond: Kimi K3 scores 93.5, compared to Claude Fable 5 at 92.6 and GPT-5.6 Sol at 94.1. This is essentially tied with the frontier.
- HLE-Full (pass@1 / pass@5): Kimi K3 at 43.5 / 56.0. Claude Fable 5 leads at 53.3 / 63.0. GPT-5.6 Sol at 44.5 / 58.0.
Coding:
- DeepSWE: Kimi K3 at 63.8, trailing Claude Fable 5 (75.4) but ahead of GPT-5.6 Sol (60.2) and well ahead of GPT-5.5 (52.5).
- FrontierMath: Kimi K3 at 31.2, competitive with GPT-5.6 Sol (32.5) and well ahead of Claude Opus 4.8 (25.0).
Frontend and Agentic Tasks:
- Arena Frontend Code: Kimi K3 ranks #1, surpassing Claude Fable 5.
- Visual QA (WorldVQA): Kimi K3 at 86.9, competitive with GPT-5.6 Sol (87.2).
The key takeaway: Kimi K3 is not quite at Claude Fable 5 or GPT-5.6 Sol level on every benchmark, but it is consistently in the same conversation. For an open-weight model that anyone can download, this is unprecedented.
The Architecture Innovations
Kimi Delta Attention (KDA)
The biggest technical story is Kimi Delta Attention. Traditional attention mechanisms compute attention scores between every pair of tokens, which costs O(n^2) (n-squared) in both compute and memory. KDA introduces a delta-based computation that achieves the same effective result with significantly lower cost.
In plain terms: earlier attention layers learn the "delta" between current token representations and store that delta efficiently. This allows the model to scale to much longer contexts without the quadratic blowup that limits models like GPT-4.
Attention Residuals (AttnRes)
Attention Residuals add a residual connection specifically around the attention computation, allowing deeper stacking of attention layers (93 layers in total) without vanishing gradient problems (where gradients become too small to update early layers in very deep networks). This is similar in spirit to how ResNet revolutionized deep vision models, but applied to transformer attention.
Together, these innovations allow Kimi K3 to maintain 1 million tokens of context while keeping inference costs manageable.
The License: Open But Not Free for Everyone
The Kimi K3 License is a custom open-weight license that is broadly permissive but has two important restrictions:
Model-as-a-Service clause: If you run a model-serving business and your aggregate revenue exceeds $20 million in any 12-month period, you need a separate commercial agreement with Moonshot AI.
Attribution for large deployments: If your product or service using Kimi K3 has more than 100 million monthly active users or generates over $20 million in monthly revenue, you must display "Kimi K3" prominently on your UI.
For individual developers, startups, and most companies, these restrictions will not apply. Internal use is completely unrestricted. The license explicitly carves out internal use from both requirements.
This is more restrictive than Apache 2.0 but significantly more open than proprietary licenses or the Llama license family. It sits somewhere between the two.
Pricing and Availability
Kimi K3 is available through multiple channels:
- Direct API: via Moonshot AI's own API at $3/million input tokens and $15/million output tokens
- OpenRouter: for pay-as-you-go access without a Moonshot account
- HuggingFace: full model weights under the Kimi K3 License
- ModelScope: alternative download mirror for users in Asia
- Third-party providers: Fireworks AI, Together, and others are expected to host inference
The pricing puts it at Claude Sonnet levels. It is significantly cheaper than Claude Opus 4.8 ($1.80 per task on the Artificial Analysis index) but more expensive than most open-weight providers.
According to Artificial Analysis, Kimi K3 costs about $0.94 per task on their long-horizon knowledge work evaluation, similar to GPT-5.6 Sol at $1.04 and roughly half the cost of Opus 4.8 at $1.80. It also uses 21% fewer output tokens than Kimi K2.6 on the same evaluation.
What This Means for Developers
For developers using AI in their workflow, the significance is clear: open-weight models have closed the gap with proprietary frontier models. Just a year ago, the best open model was clearly behind GPT-4. Today, Kimi K3 is competitive with GPT-5.5 and holds its own against GPT-5.6 Sol on several benchmarks.
This matters because open-weight models give you freedom. No vendor lock-in. No surprise API deprecations. You can switch providers, self-host eventually, and fine-tune the model yourself. For teams building AI features into their products, this is a big deal.
The Chinese AI ecosystem is now producing three distinct contenders: DeepSeek, Kimi (Moonshot AI), and Qwen (Alibaba). Each has different strengths. The "open-weight AI wars" narrative that started with Llama has evolved into a global competition with multiple power centers.
The Bottom Line
Kimi K3 is the new king of open-weight models. It is not the cheapest to run, and it does not beat Claude Fable 5 across the board. But for an open model that anyone can download, inspect, and fine-tune, competing with the best proprietary models is remarkable.
The weights are available now on HuggingFace. The tech report (PDF) is on GitHub. The API is live on Kimi.com.
The open frontier is here.
Sources:
Top comments (0)