**At midnight UTC on July 27, Beijing-based Moonshot AI released the full open weights of Kimi K3 — a 2.8-trillion-parameter mixture-of-experts model that is the largest open-weight release in history.**
The complete weights, roughly 1.4 terabytes using MXFP4 quantization, are now free to download. The model uses a mixture-of-experts architecture where approximately 50 billion parameters activate for any given token, keeping inference costs manageable despite the enormous total size. It carries a 1-million-token context window and native vision capabilities.
The performance claims have shaken the AI establishment. In blind developer testing on a leading frontend coding arena, Kimi K3 took the top spot with a score that edged past both Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol. It ranked fourth on a widely watched composite intelligence index, ahead of Claude Opus 4.8, and led results on marathon software engineering and program synthesis benchmarks.
An open Chinese model within striking distance of the frontier
Analysts urge measured interpretation. Moonshot's leaderboard figures blend results gathered through different agent harnesses, and harness choice can swing scores by 10 to 26 points on equivalent tasks. On broad, general-purpose evaluations, K3 still trails the best systems from Anthropic and OpenAI.
Yet the headline is not that K3 wins every test — it is that the distance between open and closed models has collapsed from years to months. An openly released Chinese system now sits within striking distance of the closed frontier for the first time.
Strategic implications
By publishing weights rather than gating access behind an API, Moonshot lets enterprises and governments run the model on their own hardware — directly addressing concerns about routing sensitive data through Chinese servers. Self-hosting neutralizes the data residency objection that has slowed adoption of Chinese cloud AI in Western markets.
For the open-source movement, K3 is both a milestone and a provocation. It demonstrates that frontier-scale training is no longer the exclusive domain of a few American firms. Chinese developers have achieved this despite tightened access to the most advanced chips, squeezing more performance from constrained hardware.
Every closed provider now faces a free alternative that tops at least some leaderboards, putting pressure on pricing. The 1.4-terabyte size means the immediate beneficiaries are large teams and inference providers with multi-GPU infrastructure — individual developers will likely wait for community-quantized versions.
"The open-weight tier crossed from promising to production-ready this month," noted AI analysis outlet Build Fast with AI. The release coincides with DeepSeek V4 already offering stable, cheap production-grade open models, creating a genuine two-tier open ecosystem.
