Kimi K3 Is the First Open 2.8 Trillion Parameter Model. Here Is What That Means for Your GPU Infrastructure
Moonshot AI launched Kimi K3 on July 16. It is a 2.8 trillion parameter model - the first open 3T-class model ever released. The weights ship July 27. Until then, you can use it on Kimi.com, Kimi Work, Kimi Code, and the Moonshot API.
This is not an incremental update from Kimi K2. The architecture changed. The scaling math changed. And the GPU requirements are something every infrastructure team needs to understand before the weights drop.
The Architecture: What Changed from K2 to K3
Kimi K2 was a dense transformer. Kimi K3 is a Mixture of Experts model with 896 total experts, of which 16 are active per token. This is not a small MoE - it is one of the sparsest at scale.
Two architectural innovations make this work.
Kimi Delta Attention (KDA). Standard attention computes the full attention matrix for every layer. KDA computes the delta - the difference between the current layer's attention and the previous layer's attention. The intuition: adjacent layers in deep transformers often compute similar attention patterns. Computing only the difference is cheaper and preserves more information flow across depth. The paper shows a 3x reduction in attention compute with no quality loss.
Attention Residuals (AttnRes). Each transformer block receives attention output not just from the previous block, but from several earlier blocks via learned residual connections. Think of it as skip connections across the depth dimension, similar to how ResNet skip connections work across width. This prevents information loss in very deep networks and allows Kimi K3 to use depth more efficiently.
Stable LatentMoE. MoE models are notoriously unstable during training. Expert load imbalance causes some experts to overtrain and others to undertrain. Stable LatentMoE introduces a latent router that assigns tokens to experts in a learned embedding space rather than directly from token representations. This stabilizes expert assignment and enables the extreme sparsity (16/896) without collapsing.
The combined effect is a 2.5x improvement in scaling efficiency compared to Kimi K2. The same compute budget produces roughly 2.5x more effective intelligence.
The Numbers: Parameters, Context, and Benchmarks
| Metric | Kimi K3 | Kimi K2 | GPT-5.6 Sol | Claude Fable 5 |
|---|---|---|---|---|
| Parameters | 2.8T | ~1T | Undisclosed | Undisclosed |
| Architecture | MoE (16/896) | Dense | MoE | MoE |
| Context | 1M tokens | 128K | 256K | 200K |
| Vision | Native | No | Native | Native |
| Open weights | July 27 | Yes | No | No |
| API pricing | $3/$15 per M tokens | $2/$8 | Undisclosed | Undisclosed |
Moonshot's own benchmarks place Kimi K3 between Claude Fable 5 and GPT-5.6 Sol on reasoning tasks, and slightly below both on coding tasks. For an open-weight model at 2.8T parameters, this is the new state of the art.
Importantly, Kimi K3 ships with max thinking effort by default at launch. This means it uses chain-of-thought reasoning for every query, similar to how o3 and Opus 4.5 operate. Low and high effort modes are coming in subsequent updates, which will let developers trade thinking depth for speed and cost.
What It Takes to Run Kimi K3
A 2.8T MoE model with 16 active experts does not fit in a single GPU. The math:
- Total parameters: 2.8T
- Active parameters per token: ~55B (2.8T × 16/896 ≈ 50B, plus shared layers)
- Memory for weights (FP16): ~5.6TB (2.8T × 2 bytes)
- Memory for active experts (FP16): ~110GB (55B × 2 bytes)
- KV cache for 1M context: ~200-400GB depending on precision
- Total inference memory: ~6TB minimum
This requires at minimum 8x H100 (80GB) or 8x H200 (141GB) with tensor parallelism across all GPUs, and even that is tight. Real production deployments will want 16x H200 or equivalent. The models that provide this level of intelligence at this parameter count are not hobbyist hardware projects. They are datacenter-scale deployments.
Moonshot is working with inference partners and open-source maintainers to ensure reliable rollout. Expect optimized inference backends (vLLM, SGLang, TensorRT-LLM) to add Kimi K3 support within days of the weight release.
The 1M Token Context Window
Kimi K3's 1 million token context is not just a marketing number. It is native - the model is trained from scratch with this context length, not fine-tuned onto it after the fact. This matters because fine-tuned long context models often show quality degradation beyond their original training length. Native long-context training means the model's attention patterns are optimized for the full 1M range.
1 million tokens is roughly 750,000 words, or the entire Lord of the Rings trilogy plus The Hobbit, or a medium-sized codebase of about 75,000 lines. Use cases that were impractical with 128K context (whole-codebase refactoring, multi-document legal analysis, full-transcript meeting summarization) become practical.
The downside: the KV cache for 1M tokens at this model size is enormous. Expect inference providers to charge a premium for long-context queries, or to implement sliding window attention that degrades gracefully rather than running the full 1M context on every request.
What This Means for the GPU Market
Kimi K3 being open-weight changes the GPU demand equation. Here is why.
When a frontier model is closed (GPT-5.6, Claude Fable 5), the only way to use it is through the provider's API. That concentrates GPU demand in the hands of a few hyperscaler deployments. When a frontier model is open-weight, every company can run their own instance. The GPU demand splinters across hundreds or thousands of deployments.
Kimi K3 at 2.8T parameters is not a model you run on a single cloud GPU instance. It is a model that creates demand for dedicated GPU clusters. The companies that host these clusters - bare metal GPU providers, neoclouds, colocation facilities with GPU capacity - see the demand spike. The companies that resell hyperscaler GPU capacity see their margins compress as customers buy dedicated hardware instead.
This is the same pattern that played out when Meta released Llama 3 405B. The open-weight release created net new GPU demand because every enterprise that wanted to run it needed 8x H100 bare minimum. Kimi K3 pushes that floor up to 16x H100 or 8x H200. The GPU requirements are high enough that most companies will rent dedicated hardware rather than try to squeeze it onto existing shared infrastructure.
How to Get Started
Moonshot's API is available now at api.moonshot.ai/v1 with the model ID kimi-k3. Pricing is $3 per million input tokens and $15 per million output tokens. This is roughly competitive with GPT-5.6 Terra and cheaper than Claude Fable 5.
For self-hosted deployment, the weights drop July 27. Start preparing now:
# Reserve GPU capacity. A minimal deployment needs:
# - 8x H200 (141GB each) or 16x H100 (80GB each)
# - Minimum 1TB RAM for model loading
# - High-bandwidth NVMe for KV cache offloading
# - InfiniBand or NVLink between GPUs (tensor parallel communication)
# Once weights are available:
git lfs clone https://huggingface.co/moonshotai/Kimi-K3
If you are deploying Kimi K3 on your own hardware, do not underestimate the infrastructure. This is not a weekend project on a single GPU. It is a production deployment that needs production hardware.
The Competitive Landscape
Moonshot has now held the title of largest open model for nine of the past twelve months. The progression:
- Kimi K1: January 2025, ~200B parameters
- Kimi K2: November 2025, ~1T parameters (dense)
- Kimi K2.5: March 2026, ~1.5T parameters (MoE)
- Kimi K2.6: May 2026, ~2T parameters (MoE)
- Kimi K3: July 2026, 2.8T parameters (MoE, KDA, AttnRes)
DeepSeek V4 launched this month, Zhipu GLM-5.2 is shipping, and ByteDance's Doubao is expected to announce a frontier model by August. The Chinese AI model race is producing open-weight models at a cadence the US market has not matched since Llama 3. If this continues, the most capable open models in 2027 will all come from Chinese labs.
ServerGurus GPU Cloud
We deploy dedicated GPU clusters in Hyderabad with H100 and H200 nodes on bare metal. No shared tenancy. No hypervisor tax. No capacity-block auctions. If you are planning to run Kimi K3 in production, we can provision a 16-GPU cluster within 48 hours.
Talk to us about GPU infrastructure for Kimi K3 or see bare metal GPU plans.