# Moonshot AI Releases Kimi K3 2.8T Open Model Matching Frontier Performance > The 2.8 trillion parameter model with 1 million token context ranks fourth on the Agent Arena and will open weights on July 27 after demand forced a subscription pause. *Published 2026-07-21 · By Marcus Vance* Kimi K3 is a 2.8 trillion parameter model with a 1 million token context window and native visual understanding developed by Moonshot AI. Moonshot AI released Kimi K3 on July 16, 2026. The model became available through the Kimi API Platform. Kimi K3 carries 2.8 trillion parameters. It supports a 1 million token context window. Native visual understanding forms part of its capabilities. ## What performance benchmarks does Kimi K3 achieve? Kimi K3 reached the fourth position on the Agent Arena leaderboard. The ranking places it alongside closed frontier systems. Kimi K3 matches Claude Opus 4.8 on agentic tasks. It also matches GPT-5.6 Sol on the same leaderboard. These results demonstrate parity with closed models in agent workflows. The leaderboard position reflects strong performance across multiple agent benchmarks. Developers can test agentic capabilities directly through the API. ## What technical innovations define the Kimi K3 architecture? Kimi K3 uses Kimi Delta Attention as its core attention mechanism. Kimi Delta Attention operates as a hybrid linear attention system. Attention Residuals supplement the main attention layers. The model applies increased MoE sparsity. It activates 16 experts out of 896 available experts. Stable LatentMoE contributes to training stability. These components together enable the reported efficiency gains. ## How does Kimi K3 improve scaling efficiency over prior versions? Kimi K3 delivers an approximate 2.5 times improvement in overall scaling efficiency compared to Kimi K2. The efficiency gain stems from the combination of Kimi Delta Attention, Attention Residuals, and Stable LatentMoE. Higher sparsity in the mixture of experts reduces active compute per token. The hybrid linear attention mechanism lowers memory requirements for long contexts. Residual connections help maintain gradient flow during training. ## What access options exist through the Kimi API Platform? Kimi K3 is accessible via the Kimi API with full 1 million token context support. Pricing follows a flat pay-as-you-go structure. Users can submit requests for both text and visual inputs. The API maintains the same context length as the underlying model. Existing subscribers continue to receive service during capacity constraints. ## What timeline governs the Kimi K3 rollout? - July 16, 2026: Public API launch of Kimi K3. - July 19, 2026: Pause on new subscriptions due to GPU capacity limits. - July 27, 2026: Release of full model weights under Modified MIT license. ## What caused the temporary pause on new subscriptions? Demand for Kimi K3 surged within 48 hours of the July 16 launch. GPUs operated close to full capacity during this period. Moonshot AI paused new subscriptions on July 19 to protect service quality. The pause prioritizes compute allocation for existing users. Current subscribers experienced no service interruption. > Over the past 48 hours, demand has pushed close to the limits of our current capacity. To protect the experience of existing subscribers, we're temporarily pausing new subscriptions and prioritizing compute for current members. Existing subscribed users are not affected.Moonshot AI ## What does the open-weight release mean for the AI community? Kimi K3 becomes the first open-source model in the 3-trillion-parameter class. Researchers gain direct access to weights for further experimentation. Fine-tuning on domain-specific data becomes possible after the release. The Modified MIT license permits broad commercial and research use. Open weights enable independent verification of the reported benchmark results. The release advances open-source scaling at the frontier scale. ## How might Kimi K3 influence enterprise AI strategies? Enterprises can evaluate the model through the current API before the weight release. The 1 million token context supports processing of lengthy enterprise documents. Native visual understanding allows integration of image and text workflows. Pay-as-you-go pricing reduces upfront infrastructure commitments. After July 27 organizations may deploy the model on their own hardware. The efficiency improvements lower the cost of running large-scale inference. ## What comparison exists between Kimi K3 and other frontier models? Kimi K3 compared with selected frontier models on key dimensionsFeatureKimi K3Claude Opus 4.8GPT-5.6 SolParameter Count2.8 trillionClosedClosedContext Window1 million tokensNot specifiedNot specifiedOpen WeightsYes (July 27)NoNoAgent Arena Rank4MatchedMatchedPricing ModelPay-as-you-goSubscriptionSubscription ## What next steps follow the July 27 weight release? The community will begin fine-tuning experiments immediately after the release. Independent evaluations of the 2.5 times efficiency claim will appear. New agent frameworks built on the open weights are expected. Moonshot AI may announce additional API features in subsequent updates. Hardware vendors will optimize inference stacks for the 2.8 trillion parameter scale. ## Sources 1. [Kimi K3 is a 2.8T-parameter model built on Kimi Delta Attention and Attention Residuals with a 1-million-token context window. It is the world's first open 3T-class model. The full model weights will be released by July 27, 2026.](https://www.kimi.com/blog/kimi-k3) 2. [Kimi K3 is Kimi’s most capable flagship model with 2.8 trillion parameters, built on Kimi Delta Attention (KDA) and Attention Residuals, with native visual understanding and a 1M-token context window. The full model weights will be released by July 27, 2026.](https://platform.kimi.ai/docs/guide/kimi-k3-quickstart) 3. [Moonshot AI paused new subscriptions to its Kimi K3 model on July 19, after demand pushed its GPUs close to full capacity within just 48 hours of launch.](https://finance.yahoo.com/technology/ai/articles/kimi-k3-demand-pushes-moonshot-021417183.html) --- Source: https://aiintelreport.com/frontier-models/moonshot-ai-kimi-k3-2-8t-model Index: https://aiintelreport.com/llms.txt · Full text: https://aiintelreport.com/llms-full.txt