Beijing-based Moonshot AI released Kimi K3 on July 16, 2026, a 2.8-trillion-parameter mixture-of-experts model the company calls "the world's first open 3T-class model." Launched with native vision, video input, and a 1-million-token context window, K3 is the largest open-weight language model ever announced — roughly 75 percent bigger than DeepSeek's V4 Pro — and Moonshot says the full weights will be freely downloadable by July 27.
Key Highlights
- 2.8 trillion total parameters in a sparse mixture-of-experts design that activates 16 of 896 experts per token
- 1-million-token context window with native vision and video understanding
- Third place overall on the GDPval-AA v2 real-world work benchmark, behind only Claude Fable 5 Max and GPT-5.6 Sol Max
- API pricing of $0.30 per million cached input tokens, $3 for cache misses, and $15 per million output tokens
- Full open weights promised by July 27, 2026, under a free open-weight distribution
Details
Kimi K3 builds on two architectural changes Moonshot published earlier this year: Kimi Delta Attention (KDA), a hybrid linear attention mechanism the company says delivers up to 6.3x faster decoding in million-token contexts, and Attention Residuals (AttnRes), which retrieves representations across model depth and adds roughly 25 percent training efficiency at under 2 percent extra cost. Combined with a set of sparsity techniques — Stable LatentMoE, Quantile Balancing, and Per-Head Muon among them — Moonshot claims an approximate 2.5x improvement in overall scaling efficiency compared with its previous K2 generation.
The benchmark story is what sets K3 apart from previous open releases. On GDPval-AA v2, which measures performance on real-world tasks across 44 occupations and 9 major industries, K3 scored 1,687 — third overall, behind Anthropic's Claude Fable 5 Max at 1,815 and OpenAI's GPT-5.6 Sol Max at 1,747.8, but ahead of Claude Opus 4.8 at 1,600. With maximum reasoning effort, the model posts 93.5 on GPQA-Diamond, 91.2 on BrowseComp, and 42.0 on the long-horizon SWE Marathon coding benchmark.
In its announcement, Moonshot was measured about the model's standing, noting that K3 "still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol," while consistently outperforming every other system it tested.
Two variants shipped at launch: K3 Max for chat and agent tasks, and K3 Swarm Max for large-scale parallel processing. The model is live on Kimi.com, the Kimi Work desktop app, the Kimi Code terminal tool, and the Kimi API platform, which is compatible with the OpenAI SDK.
Impact
K3 collapses the gap between open-weight and frontier proprietary models to a few months at most, and it does so at aggressive pricing. Cached input at $0.30 per million tokens — with Moonshot reporting cache-hit rates above 90 percent in coding workloads — undercuts comparable proprietary tiers by a wide margin, a direct challenge to Anthropic and OpenAI in the fast-growing agentic coding market.
For developers and businesses in Tunisia and the wider MENA region, an open 3T-class model changes the self-hosting calculus. The weights ship in MXFP4 quantization with MXFP8 activations, and Moonshot recommends a supernode of 64 or more accelerators for full deployment — beyond most startups, but within reach of regional cloud providers, research institutions, and sovereign AI initiatives that want frontier-class capability without dependence on US API providers.
Background
Moonshot AI has iterated rapidly through 2026, following Kimi K2.6 — its open-source coding agent swarm release — and the K2.7 Code model earlier this year. The company also published the Attention Residuals research that underpins K3's architecture. The release lands amid an intense open-versus-closed contest: Zhipu's GLM-5, DeepSeek's V4, and Ant Group's trillion-parameter Ling 2.6 have all pushed open weights toward the frontier in recent months, while Chinese labs face US export controls on advanced training hardware.
The launch also came one day before the World Artificial Intelligence Conference opened in Shanghai, sharpening the geopolitical subtext of China's biggest open-model release to date.
What's Next
The full open weights are scheduled for release by July 27, which will confirm the license terms and let independent researchers verify Moonshot's benchmark claims. Attention now turns to how proprietary leaders respond on pricing — and to whether inference providers outside China pick up K3 hosting at scale once the weights land.
Source: Moonshot AI