Kimi K3 is the largest open-weight model released to date. Moonshot AI announced it on July 16, 2026, and published the full weights on July 27, making it 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.
The hardware requirement is the catch. At 2.8T total parameters, Kimi K3 requires roughly 1.4 TB of VRAM just to load the weights at native precision. Running it privately requires multi-node datacenter infrastructure that most teams don't have or can't afford.
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 enables up to 6.3x faster decoding in million-token contexts.
- Attention Residuals (AttnRes), that improves training efficiency at long sequence lengths.
The MoE layer uses a Stable LatentMoE framework that activates 16 of 896 experts per forward pass, yielding roughly 2.5x better scaling efficiency than Kimi K2. Training used quantization-aware training (QAT) from the supervised fine-tuning stage onward, producing native 4-bit weights (MXFP4) rather than applying quantization post-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 performs at frontier level while trailing the top proprietary models. On GDPval-AA v2, a benchmark measuring performance across 44 occupations and 9 industries, K3 scored 1,687, placing third behind Claude Fable 5 Max (1,815) and GPT-5.6 Sol Max (1,747.8), and ahead of Claude Opus 4.8 (1,600).
| Benchmark | Kimi K3 | Claude Fable 5 Max | GPT-5.6 Sol Max | Claude Opus 4.8 |
|---|---|---|---|---|
| GDPval-AA v2 | 1,687 | 1,815 | 1,747.8 | 1,600 |
| AA-Briefcase (agentic) | 1,527 | 1,587 | 1,495 | n/a |
| BrowseComp | 91.2 | n/a | n/a | n/a |
| Benchmark figures from Moonshot AI's technical report and third-party evaluations published July 2026. K3 results use max thinking effort and temperature 1.0. |
||||
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 MIT-like 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.
The Memory Reality
Kimi K3's native 4-bit (MXFP4) weights occupy roughly 1.4 TB at 2.8T parameters, with a full repository footprint of approximately 1.56 TB. A single session adds around 15 GB of KV cache state plus roughly 60 GB for activations and runtime overhead, bringing the single-user total to around 1.63 TB. Serving multiple concurrent users scales that figure higher.
Serving K3 at native precision is also a hardware generation decision. A100, H100 and H200 can load MXFP4 weights but will dequantize them to a higher precision at runtime. This would effectively preserve the memory savings but lose the throughput benefits. Native MXFP4 execution requires Blackwell-generation hardware: NVIDIA B200/B300 or AMD MI350X/MI355X.
The most compressed available build is Unsloth's 1-bit Dynamic GGUF, which brings weight size down to ~594 GB and requires ~650 GB of total memory to run. Unsloth's documented workaround is a 512 GB Mac Studio linked to a 128 GB node, pooling 640 GB of unified memory. An 8x H100 node (640 GB VRAM) with sufficient system RAM can also load it via llama.cpp offloading, but inference is impractical. To avoid experimental workarounds and get practical throughput, the 1-bit build needs a minimum of 4x B200 (768 GB VRAM) with native MXFP4 support.
| Build | Weight Size | Memory Floor | Minimum for Practical Inference |
|---|---|---|---|
| Unsloth 1-bit dynamic | ~594 GB | ~650 GB | 4x B200; 8x H100 (experimental, impractical)1 |
| Unsloth 2-bit dynamic | ~861 GB | ~900 GB+ | 8x H200 |
| Native 4-bit (MXFP4) | ~1.4 TB | ~1.5 TB (single user) | 8x B200 |
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 have K3 available 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 August 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.
If your team has data-sovereignty requirements, needs high sustained throughput, or wants to fine-tune on custom weights, self-hosting an earlier Kimi model is worth evaluating. Kimi K2.6 is a 1T-parameter MoE model with 32B active parameters that runs on an 8x GPU cluster. K2.7 Code, released in June 2026, targets agentic coding and uses roughly 30% fewer thinking tokens than K2.6 for equivalent task quality.
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.
See Thunder Compute's current GPU availability and pricing for on-demand A100 and H100 instances.
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 |
| Multimodal | Text + vision | Text + vision + video |
| Min. self-host memory | ~247 GB (1.8-bit) | ~650 GB (1-bit, experimental) |
| Realistic self-host floor | 8x A100 / H100 (640 GB) | 4x B200 (1-bit); 8x B200 (native 4-bit) |
| License | Modified MIT | Kimi K3 License |
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 run at native precision.
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 fast: Unsloth published the 1-bit Dynamic GGUF just two days after the weights landed, and it represents the floor of what's practically quantizable without significant quality loss. The hardware requirement for K3 itself is unlikely to drop much further. Broader access will come from distilled derivatives, smaller models trained on K3's outputs, rather than from K3 itself getting cheaper to run.
For open-weight Kimi models that run on multi-GPU hardware today, the K2 guide covers the full self-hosting path. 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 Kimi K3
Kimi K3 is the first open-weight model in the 3T-parameter class, downloadable under a permissive license with frontier-level benchmark performance. For practical access today, the Kimi API and OpenRouter are the right starting points. Self-hosting K3 at native precision requires Blackwell or MI350X-class infrastructure; H100 clusters can load quantized builds but dequantize at runtime and produce impractical throughput. For teams that want a self-hosted open-weight Kimi model on realistic GPU hardware now, the K2 family is the path.
FAQ
What Is Kimi K3?
Kimi K3 is an open-weight multimodal model from Moonshot AI with 2.8T total parameters and 104B active parameters per forward pass. It supports text, images, and video natively, ships with a 1M-token context window, and is the largest open-weight model released to date. Weights were published July 27, 2026.
When Was Kimi K3 Released?
Moonshot AI announced Kimi K3 on July 16, 2026, with immediate API access. Full open-weight release, including model weights and the technical report, followed on July 27, 2026.
Is Kimi K3 Open Source?
Kimi K3 is open-weight, not fully open source. The trained weights are freely downloadable under the Kimi K3 License, which is MIT-like for most uses. Two commercial gates apply: MaaS providers above $20M annual revenue need a separate Moonshot agreement, and products above 100M MAU must display prominent K3 attribution.
Can You Run Kimi K3 Locally?
Not practically. The 1-bit build (~594 GB) can load on an 8x H100 cluster with system RAM or a 512 GB Mac Studio linked to a 128 GB node, but both produce impractical throughput. H100 and older GPUs also dequantize MXFP4 at runtime, losing K3's native speed benefits. The minimum for practical inference is 4x B200 (1-bit build) or 8x B200 (native 4-bit).
What Changed Between Kimi K2 and Kimi K3?
Kimi K3 triples total parameters (1T to 2.8T), increases active parameters from 32B to 104B, extends context from 256K to 1M tokens, adds video modality, and introduces the bespoke Kimi K3 License in place of Modified MIT. It also requires Blackwell (B200/B300) or MI350X-class hardware for native precision; H100 and H200 dequantize at runtime.