"Moonshot AI released the model weights and technical report for Kimi K3, a 2.8T MoE model with native vision and a 1M-token context window."
Reliable sources confirm that Moonshot AI has released the model weights and technical report for Kimi K3, a 2.8-trillion-parameter MoE model with native visual understanding and a 1-million-token context window.
Why this score
↑ Official release of Kimi K3 model weights and technical report verified on Hugging Face and tech blogs.
↑ Model specifications match a 2.8T Mixture-of-Experts architecture with 1M-token context and native vision.
↑ Architectural focus on efficiency gains (such as Kimi Delta Attention) aligns with the claim.
🌐 FULL SPECTRUM SETTLED FACT
ConsensusMoonshot AI did release the Kimi K3 open weights and technical report (around 26–27 July 2026), confirming a 2.8T-parameter MoE model (~104B active) with native multimodal/vision support and a 1M-token context window; the new KDA + Attention Residuals + Stable LatentMoE architecture is stated to deliver ~2.5× scaling efficiency versus Kimi K2.
↔ The release is not meaningfully 'open' and the efficiency/performance claims are overstated marketing.Who: Open-source purists, license analysts, some Western AI commentators and X users
Strongest argument: Weights are public on Hugging Face but under a modified/non-OSI license with commercial/revenue-share restrictions; full training data and recipes are incomplete; independent runs show it still trails top proprietary models (Claude Fable 5, GPT-5.6 Sol) overall and the 2.5× figure is self-reported without broad third-party compute-normalized verification; the model is too large (~1.5 TB+) for prac
Assessment: License caveats and size barriers are well-documented and reduce real-world openness; trailing proprietary SOTA is acknowledged even by Moonshot; the 2.5× efficiency claim rests primarily on their tech report and has not been comprehensively refuted or independently replicated at scale.
↔ Kimi K3's capabilities may derive substantially from distillation/IP misuse rather than pure novel scaling.Who: US officials/allegations, some Western tech-policy outlets and commentators
Strongest argument: Reports claim Moonshot used large-scale querying of Anthropic (and possibly other) models via fake accounts to distill frontier capabilities, raising questions whether the architecture gains and benchmark results reflect genuine innovation or appropriated outputs, especially under US export controls on advanced chips.
Assessment: Allegations circulate and fit broader US–China AI friction narratives, but public primary evidence (court filings, definitive forensic proof) remains limited; the released tech report and architecture details provide a counter-narrative of original research that has not been conclusively disproven.
Open questions• Independent, compute-normalized verification of the claimed 2.5× scaling-efficiency gain and exact active-parameter behavior across long-horizon agentic tasks.
• Full enforceability and practical impact of the modified license on commercial fine-tunes, distillations, and third-party hosting.
• Extent to which any distillation from proprietary models contributed to final capabilities versus the published architectural and data innovations.