On July 27, 2026, Moonshot AI published the full weights of Kimi K3, a natively multimodal mixture-of-experts (MoE) model with 2.8 trillion total parameters, 104 billion active parameters and a 1-million-token context window. Moonshot calls it "the world's first open 3T-class model," and at release it was the largest open-weight model available. Kimi K3 is aimed at long-horizon coding and agentic knowledge work, and its license introduces conditions for large commercial users that developers need to read before building on it.

Key Facts#

  • Release: Moonshot announced Kimi K3 in mid-July 2026 and released the full weights on July 27, with a GitHub repository, a technical report and downloads on Hugging Face.
  • Scale: 2.8T total parameters with 104B activated, using 896 experts, of which 16 are selected per token, plus 2 shared experts.
  • Architecture: 93 layers that mix 69 Kimi Delta Attention (KDA) layers with 24 gated multi-head latent attention layers, plus Attention Residuals (AttnRes), a design Moonshot published earlier in 2026, and a Stable LatentMoE framework.
  • Context and modalities: a 1,048,576-token context window, with text, image and video understanding through a 401M-parameter vision encoder.
  • Quantization: trained with quantization awareness from the fine-tuning stage onward, and shipped with MXFP4 weights and MXFP8 activations.
  • API: available as kimi-k3 on Moonshot's platform through OpenAI- and Anthropic-compatible APIs. Self-hosting is supported on vLLM, SGLang and TokenSpeed.
  • License: the Kimi K3 License, a modified MIT-style license with conditions for large model-hosting businesses and for very large consumer products.

What Happened#

Moonshot released Kimi K3 as a single model rather than a family of sizes. It says the Stable LatentMoE framework, which activates 16 of 896 experts, delivers about 2.5 times better overall scaling efficiency than Kimi K2. Moonshot positions the model for long engineering sessions in large repositories, terminal-driven work such as GPU kernel optimization, and knowledge work that produces research reports, dashboards and visualizations.

Moonshot's evaluation table compares Kimi K3 at maximum reasoning effort with Claude Fable 5, GPT-5.6 Sol, Claude Opus 4.8, GPT-5.5 and Z.ai's GLM-5.2. The results are mixed in an informative way. On GPQA Diamond, K3 scored 93.5, compared with 94.1 for GPT-5.6 Sol and 92.6 for Claude Fable 5, and Moonshot reports strong scores on agentic coding and tool-use benchmarks. Its footnotes also show how hard these comparisons are. Different models ran in different agent harnesses, some tests used H20 GPUs instead of the official hardware, and Moonshot notes that Claude Fable 5 hit fallbacks on a significant share of tasks in some runs. Treat all of these as vendor-reported results.

Community interest was immediate. According to press reports, Hugging Face CEO Clem Delangue said Kimi K3 reached the top of the platform's trending chart within 30 minutes of release, the fastest rise the site had recorded.

Background#

Moonshot's flagship open-weight series began with Kimi K2 in July 2025, a 1T-parameter agentic model, and continued with the multimodal Kimi K2.5 in January 2026. K3 nearly triples the total parameter count and more than triples the active parameters. It also replaces about three-quarters of its attention layers with Kimi Delta Attention, a linear-attention mechanism that Moonshot introduced in its open Kimi Linear research in 2025.

The license is the biggest change from K2. K2's modified MIT license only required very large products to display the model name. The Kimi K3 License keeps a similar rule for products with more than 100 million monthly active users or more than $20 million in monthly revenue, but adds a new condition: a company that runs a "Model as a Service" business with more than $20 million in revenue over any 12 months must sign a separate agreement with Moonshot before using the model commercially. Internal use and access through Moonshot's own products or certified partners are exempt. Hugging Face's summer 2026 report flagged this as part of a shift, with the largest new Chinese releases starting to add revenue-share or non-commercial conditions after a year of mostly Apache 2.0 and MIT licenses.

Why It Matters for Developers#

For most teams, Kimi K3 is an API model with an unusual escape hatch. Even with native MXFP4 weights, 2.8T parameters come to well over a terabyte, so self-hosting means a multi-GPU, often multi-node, cluster. The value of open weights here is optionality: you can move to self-hosting or a certified partner later, or run the model in a sovereign environment, without changing models.

The integration details matter more than usual. Kimi K3 always thinks, and returns its reasoning in a reasoning_content field. Effort is set with a top-level reasoning_effort field that accepts low, high or max, and the default is max, which is the most expensive setting. Most important, Moonshot says K3 was trained in a "preserved thinking" mode: for multi-turn conversations and tool calls, you must send the complete assistant message back, including reasoning_content and tool_calls, not just the visible content. If your .NET client, Microsoft.Extensions.AI middleware or chat-history store drops that non-standard field, quality can degrade without any error. Test the round trip explicitly, and set the effort level deliberately per scenario to control cost and latency, using the practices in our LLM observability and cost guide.

Check the license against your business model. If you are an internal developer building on Moonshot's API, the new terms are unlikely to affect you. If you operate a platform that resells model access, the Model-as-a-Service clause may require a direct agreement with Moonshot.

Finally, compare agent frameworks, not only models. Moonshot recommends its own Kimi Code CLI as the harness for K3, and its benchmark notes show that scores shift with the harness. When you evaluate K3 for coding agents or AI-assisted development, run it in the harness you will actually use.

What's Next#

Independent testing has started. In one public but unofficial evaluation of the IMO 2026 problems, graded by verifier agents rather than the IMO, Kimi K3 reached a verified 42 out of 42 after reviewer-feedback repair rounds, up from 36 on its first pass. More third-party benchmarks, quantized builds and fine-tunes will follow. Hugging Face's summer report notes that GGUF builds of the roughly 2.8T model already existed in July, although running them still takes substantial hardware.

The broader question is licensing. If other frontier open-weight labs follow Moonshot and Alibaba in adding commercial conditions to their largest models, "open weights" will increasingly come with terms that developers must track as carefully as API pricing. Our guide to AI agent architecture patterns covers how to keep model choice swappable when those terms change.

Sources#