Moonshot AI Unveils Kimi K3: A 2.8-Trillion Parameter Multimodal Model

Moonshot AI has released Kimi K3, a 2.8-trillion-parameter open-weight multimodal model, as reported by VentureBeat. The Beijing-based startup's flagship ships a 1-million-token context window, native visual understanding, and two internally developed mechanisms — Kimi Delta Attention and Attention Residuals — that the company claims deliver consistent scaling gains. Full model weights are scheduled to drop on July 27, timed ahead of the 2026 World Artificial Intelligence Conference in Shanghai.

Architectural claims and their auditability

The headline figure of 2.8T parameters places Kimi K3 roughly 75 percent above DeepSeek's V4 Pro, though total parameter counts remain a weak proxy for usable compute absent a published active-parameter count per token. Two components anchor the architectural story: Kimi Delta Attention, described as a hybrid linear attention mechanism, and Attention Residuals, framed as a drop-in replacement for standard residual connections. Both were previously published as open research on GitHub, which permits a methodological audit ahead of the weights release. The API surface is OpenAI-SDK-compatible, and an "always-on" reasoning mode ships by default. Pricing lands at $3 per million input tokens and $15 per million output tokens, with cached input at $0.30 per million — mid-tier Western pricing at claimed frontier-tier capability.

Benchmark posture versus independent reproduction

Reported scores concentrate on two evaluations. On GDPval-AA v2 — measuring real-world tasks across 44 occupations and 9 industries — Kimi K3 reaches 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). On AA-Briefcase, a private agentic benchmark from Artificial Analysis, the model scores 1,527, trailing Fable 5 Max (1,587) and beating GPT-5.6 Sol Max (1,495). The release also cites state-of-the-art results on long-horizon coding and web browsing benchmarks. All three claims rest on vendor-curated tables and one private harness pending independent reproduction.

Verification checklist once weights land

The decisive date is July 27. Until then, every reported number is an unverifiable claim against a closed checkpoint; the live API at kimi.com is the only tractable artifact. Practitioners should track three points: the active-parameter count per forward pass (since total parameters do not equal FLOPs per inference), reproducibility of the GDPval-AA v2 and AA-Briefcase deltas under independent harnesses, and parity between the API and the eventual open-weight release. The same density questions reappear in Alibaba's parallel Qwen3.8-Max-Preview — a 2.4T-parameter multimodal model previewed via Token Plan subscription, as covered by MarkTechPost — where the unresolved active-parameter count and serving arithmetic on commodity hardware will determine practical deployability. Frontier-scale open-weight releases are now arriving faster than reproducibility audits can clear them; the next two weeks will separate benchmark claims from engineering reality. Resource-budgeting trade-offs of this magnitude surface in any domain concerned with maximizing output per unit input, from model serving to dietary optimization.