Kimi K3 is Moonshot AI's 2.8T-parameter open-weight model, released July 27, 2026, and the first open model in the 3T-parameter class. It sits just below the top proprietary models on most benchmarks and is freely downloadable for research and many commercial uses.
Running Kimi K3 takes datacenter-class hardware: the model occupies roughly 1.56 TB of memory in its native MXFP4 (4-bit) format, well beyond any single GPU. For most teams, a hosted API is the only option.
This guide covers the hardware requirements at a high level, the local and hosted options, and what to run instead when K3's footprint is out of reach.
Takeaways
- No single GPU runs Kimi K3. In its native MXFP4 (4-bit) format the model needs roughly 1.56 TB of memory, requiring multi-node GPU clusters.
- Native-speed inference requires Blackwell. MXFP4 execution runs on NVIDIA B200/B300 or AMD MI350X; H100 and H200 load the weights but dequantize at runtime.
- A hosted API is the practical path. The Moonshot API and OpenRouter serve K3 at $3/$15 per 1M tokens.
- Kimi K2.6 and K2.7 are the self-hostable alternatives. Both run on an A100 or H100 node, available on Thunder Compute from $1.09/hr.
- Kimi K3 leads open-weight models on independent benchmarks, scoring 57 on the Artificial Analysis Intelligence Index, first among open-weight models.
What Is Kimi K3?
Kimi K3 is Moonshot AI's flagship model, succeeding the Kimi K2 family. It uses a Mixture-of-Experts (MoE) architecture with 2.8T total parameters and 104B active parameters per forward pass. The model is natively multimodal, supporting text, images, and video in a single architecture, and ships with a 1M-token context window.
The architecture introduces two new components:
- Kimi Delta Attention (KDA), a hybrid linear attention mechanism that speeds up decoding in million-token contexts.
- Attention Residuals (AttnRes), which improve training efficiency at long sequence lengths.
The MoE layer uses a Stable Latent MoE framework that activates 16 of 896 experts per forward pass. Moonshot reports this improves scaling efficiency over Kimi K2 by roughly 2.5x. Training used quantization-aware training (QAT) from the supervised fine-tuning stage onward, producing native 4-bit weights (MXFP4) rather than applying quantization after training.
Kimi K3 Release Date
Kimi K3 was announced on July 16, 2026, with immediate access through the Kimi web app, Kimi Work, Kimi Code, and the Kimi API. The full open-weight release followed on July 27, 2026, when the model weights, technical report, and a KDA prefill-cache implementation for vLLM were published together. Moonshot coordinated with inference partners including NVIDIA, AMD, and the vLLM and llama.cpp maintainers to prepare day-zero serving support.
Kimi K3 Benchmarks vs. Claude and GPT
Kimi K3 is the strongest open-weight model on independent evaluations, while trailing the top proprietary models.
| Benchmark | Kimi K3 | Claude Fable 5 Max | GPT-5.6 Sol Max | Claude Opus 4.8 |
|---|---|---|---|---|
| GDPval-AA v21 | 1,668 | 1,760 | 1,743 | 1,600 |
| AA-Briefcase1 | 1,548 | 1,583 | 1,495 | 1,347 |
| BrowseComp2 | 91.2 | 88.0 | n/a | n/a |
| Leaderboard Elos shift as models and reasoning-effort tiers are added, so treat these as directional. Figures reflect max reasoning effort. 1Elos from Artificial Analysis (July 2026 snapshot); 2 From Moonshot AI's technical report. |
||||
K3 leads all tested open models across coding, knowledge work, and long-context reasoning. Its agentic performance is especially strong: K3 sustains long engineering sessions with minimal supervision, navigates large codebases, and orchestrates terminal tools.
Is Kimi K3 Open Source?
Kimi K3 is open-weight, not fully open source in the OSI sense. The trained weights are freely downloadable, but the training data and pipeline are not publicly released. For most research, internal deployment, and commercial integration, the open-weight release is sufficient.
The Kimi K3 License
Kimi K3 ships under a bespoke Kimi K3 License, not the Modified MIT license used by Kimi K2. The license is permissive for most uses but includes two commercial gates:
- Organizations hosting K3 as a paid model-as-a-service with annual revenue above $20M need a separate agreement with Moonshot AI.
- Products exceeding 100M monthly active users must display prominent Kimi K3 attribution.
For most developers and teams, those thresholds are not relevant. Research use, internal deployment, fine-tuning, and commercial integration below those thresholds are all permitted. Read the LICENSE file in the official Hugging Face repository before productizing, and don't infer terms from the earlier K2 releases; the licenses differ.
For more information, consult the LICENSE file in the official GitHub repository.
Can You Run Kimi K3 Locally?
Running Kimi K3 locally is not feasible on consumer hardware, workstations, or most cloud GPU tiers. The memory requirements are well outside what any single-node setup can satisfy.
How Many GPUs Does Kimi K3 Need?
Kimi K3 cannot run on a single GPU; it requires a multi-node GPU cluster. In fact, Moonshot recommends a supernode configuration of 64 or more accelerators for production serving, reflecting the communication load of routing 16 of 896 experts per token.
Community-quantized builds reduce the footprint but still require several datacenter-class GPUs and only reach practical throughput with Blackwell-class hardware.
The Memory Reality
Kimi K3's native 4-bit (MXFP4) weights occupy roughly 1.4 TB of memory at 2.8T parameters, before runtime overhead. That footprint alone rules out any single-node deployment. MXFP4 is the format Moonshot trained and released; there is no separate BF16 checkpoint to upcast to, which at 2.8T parameters would need roughly 5.6 TB.
Serving K3 in its native MXFP4 format is also a hardware-generation decision. A100, H100, and H200 can load MXFP4 weights but dequantize them to a higher precision at runtime, which preserves the memory savings but loses the throughput benefits. Native MXFP4 execution requires Blackwell-generation hardware: NVIDIA B200/B300 or AMD MI350X/MI355X.
The most compressed community build, Unsloth's 1-bit Dynamic GGUF, reduces the footprint substantially but still needs requires ~650 GB of memory across multiple datacenter GPUs, without production-level throughput or quality.
What About Ollama's kimi-k3:cloud Tag?
Ollama's kimi-k3:cloud tag is not local inference. The cloud tag forwards prompts to Moonshot's infrastructure via Ollama's servers, requires a paid Ollama account that consumes usage credits, and stores only a small manifest file on your machine. Prompts leave your device the same way they do with any API call.
True local inference in Ollama and LM Studio depends on llama.cpp support that was still in progress at the time of the weight release. Check the llama.cpp issue tracker for current status before assuming support has landed.
For a broader look at what Ollama's cloud tags do and how they differ from running a model locally, see the Thunder Compute guide to Ollama.
How to Access Kimi K3 Today
The practical access path for most teams is a hosted API. Several providers serve K3 with no infrastructure setup required.
| Provider | Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|---|
| Moonshot AI API | kimi-k3 | $3.00 | $15.00 |
| OpenRouter | moonshotai/kimi-k3 | $3.00 | $15.00 |
| Ollama (cloud tag) | kimi-k3:cloud | Usage credits | Usage credits |
| Pricing as of September 2026. Subject to change. Verify on provider billing pages before committing to a workflow. | |||
OpenRouter and the Moonshot API are the right starting points for most developers. Both provide a stable endpoint, token-based pricing, and no infrastructure overhead. Ollama's cloud tag works if you're already in the Ollama ecosystem and data residency is not a requirement.
When It Makes Sense to Self-Host a Kimi Model
Self-hosting Kimi K3 is only practical for organizations that already operate Blackwell or MI350X-class multi-node clusters. For everyone else, the API is the correct answer.
Self-hosting an earlier Kimi model is the realistic path if your team has data-sovereignty requirements, needs high sustained throughput, or wants to fine-tune on custom weights. Kimi K2.6 is a 1T-parameter MoE model with 32B active parameters that runs on a single multi-GPU node. K2.7 Code, released in June 2026, targets agentic coding and is more token-efficient than K2.6 for equivalent task quality.
Both models run on hardware you can rent on demand today. Thunder Compute offers A100 80GB at $1.09/hr and H100 PCIe at $3.20/hr, in 1x to 8x configurations with per-minute billing and no egress fees. An A100 or H100 node fits a self-hosted K2.6 or K2.7 deployment, which makes it the practical entry point for an open-weight Kimi model on your own infrastructure.
The Thunder Compute guide to running Kimi K2 with Ollama covers the full setup, recommended multi-GPU configurations, and GPU-hour pricing from a 4x A100 to an 8x H100 node.
Kimi K3 vs. Kimi K2: What Changed
Kimi K3 is not an incremental update to Kimi K2. The two models share the Kimi brand and the open-weight philosophy, but their architecture, parameter count, and serving requirements are in different classes.
| Attribute | Kimi K2.6 | Kimi K3 |
|---|---|---|
| Total parameters | 1T | 2.8T |
| Active parameters | 32B | 104B |
| Context window | 256K tokens | 1M tokens |
| Native precision | BF16 / INT4 | MXFP4 (QAT) |
| Native precision hardware | Hopper / Ampere | Blackwell (B200/B300) or MI350X/MI355X |
| Minimum serving cluster | 8x A100 / H100 | 16x B200 or 8x B3001 |
| Multimodal | Text + vision | Text + vision + video |
| License | Modified MIT | Kimi K3 License |
| 1 K3's ~1.56 TB checkpoint exceeds an 8x B200 node (1,440 GB), so it cannot load there. Native 4-bit serving starts at roughly 16x B200 across two nodes, or an 8x B300 node. | ||
The MoE sparsity increase is the key architectural difference between K2 and K3. Kimi K2 activates 32B of 1T parameters per token; Kimi K3 activates 104B of 2.8T, using a different expert routing structure (16 of 896 experts) that Moonshot AI reports yields 2.5x better scaling efficiency. K3 is meaningfully more capable per FLOP, but the full weight set is roughly 3x heavier and requires proportionally more serving infrastructure, plus a newer GPU generation to serve its native 4-bit format at full speed.
What the Open-Weight Release Means for the Community
The open-weight release of Kimi K3 enables work that API access alone does not. The community can now distill smaller task-specific models from K3, fine-tune for domain-specific use cases, and reproduce or extend Moonshot's architectural experiments (KDA, AttnRes, Stable LatentMoE) in open research.
The community moved quickly: Unsloth published a 1-bit Dynamic GGUF within days of the weight release. The hardware requirement for K3 itself is unlikely to drop much further, so broader access will come from distilled versions, smaller models trained on K3's outputs, rather than from K3 getting cheaper to run.
For quantized local inference on smaller models, see the Thunder Compute guide to LM Studio and the guide to running Unsloth for fine-tuning and GGUF export workflows.
Last Thoughts on How to Run Kimi K3
Kimi K3 is the first open-weight model in the 3T-parameter class, but running it takes datacenter-class infrastructure. For practical access today, the Kimi API and OpenRouter are the right starting points. Self-hosting K3 at native speed requires Blackwell or MI350X-class hardware.
For teams that want a self-hosted open-weight Kimi model on GPU hardware they can rent now, Kimi K2.6 or K2.7 on an A100 or H100 node is the realistic path.
FAQ
Can you run Kimi K3 locally?
Not practically. Kimi K3 requires a multi-node datacenter GPU cluster; no single GPU or workstation can load it. Native-precision inference needs Blackwell-class hardware (B200/B300 or MI350X), and even the most compressed community builds need multiple datacenter GPUs without practical throughput.
What is Kimi K3?
Kimi K3 is Moonshot AI's open-weight multimodal model with 2.8T total parameters and 104B active parameters per forward pass. It supports text, images, and video, ships a 1M-token context window, and is the largest open-weight model released to date.
When was Kimi K3 released?
Moonshot AI announced Kimi K3 on July 16, 2026, with immediate API access. The full open-weight release, including model weights and the technical report, followed on July 27, 2026.
What can I self-host instead of Kimi K3?
Kimi K2.6 and K2.7 Code are the realistic self-hosted Kimi models. Both use a 1T-parameter MoE with 32B active parameters that fits on an A100 or H100 cluster, available on Thunder Compute from $1.09/hr.
How does Kimi K3 compare to Claude and GPT?
Kimi K3 is the strongest open-weight model on independent benchmarks but trails the top proprietary models. It scores 57 on the Artificial Analysis Intelligence Index, first among open-weight models, and ranks second on AA-Briefcase behind Claude Fable 5.
What changed between Kimi K2 and Kimi K3?
Kimi K3 triples total parameters from 1T to 2.8T, raises active parameters from 32B to 104B, extends context from 256K to 1M tokens, and adds video input. It moves to the bespoke Kimi K3 License and needs Blackwell-class hardware for native MXFP4 execution.