China's Kimi K3 Goes Fully Open This Week — A 2.8-Trillion-Parameter Model
Moonshot AI is set to release Kimi K3's full weights, making the largest open-weight model ever built publicly downloadable — and it already beats Claude Opus 4.8 on key benchmarks.
Chinese lab Moonshot AI is set to release the full weights of Kimi K3 — a 2.8-trillion-parameter model that Moonshot calls the world's first "open 3T-class" system: large enough to compete with the best proprietary models on the market, while still being free for anyone to download and run.
Two launches, two dates
Kimi K3 is arriving in two stages. The model went live via API on July 16, and the full open weights — under a Modified MIT license — are scheduled to land July 27 at huggingface.co/moonshotai. Expect roughly a 594GB download for the native MXFP4 safetensors release once it lands.
What's under the hood
Kimi K3 is a mixture-of-experts model built on Kimi Delta Attention (KDA), a hybrid linear attention mechanism, paired with a technique Moonshot calls Attention Residuals. It ships with native vision support and a one-million-token context window.
How it actually performs
On GDPval-AA v2 — a benchmark measuring real-world task performance across 44 occupations and 9 industries — Kimi K3 scored 1,687, placing it third overall:
| Model | Score | |---|---| | Claude Fable 5 Max | 1,815 | | GPT-5.6 Sol Max | 1,747.8 | | Kimi K3 | 1,687 | | Claude Opus 4.8 | 1,600 |
Kimi K3 also beat Claude Opus 4.8 and GPT-5.5 on coding and general-agent benchmarks, and debuted at #1 on Arena's Frontend Code leaderboard — a 17-place jump over its predecessor, Kimi K2.6, landing ahead of Claude Fable 5 on that specific test.
"Open-weight" means the trained model's parameters are published for anyone to download and run themselves — unlike closed models such as GPT-5.6 or Claude, which are only accessible through the company's own API or apps.
Why it matters
For developers and researchers: a model this capable, with weights this open, means serious frontier-level AI capability no longer requires an API relationship with a U.S. lab — it can run on your own infrastructure.
For the U.S.-China AI race: Moonshot built a model competitive with Claude and GPT-5.6 while working within U.S. compute export restrictions, undercutting the assumption that those controls would keep a meaningful capability gap in place.
For enterprises: an open-weight model beating Claude Opus 4.8 and GPT-5.5 on real benchmarks gives cost-sensitive teams a credible, self-hostable alternative to closed frontier APIs for the first time at this capability tier.
What to watch
- How Kimi K3 performs on independent benchmarks once the community can test the open weights directly, beyond Moonshot's own reported numbers
- Whether OpenAI or Anthropic respond with their own open-weight releases
- Adoption speed — a 594GB download is a real barrier outside well-resourced labs and enterprises
Sources: VentureBeat, CNBC, Tom's Hardware
NextGen AI Digest Editorial
Editorial Team
Reporting and analysis from the NextGen AI Digest newsroom — covering AI, agentic systems, SaaS, and the future of technology. Every piece is factual, sourced, and cited. Built and published by the team at Peaders.
Keep reading
OpenAI Releases GPT-5.6 to the Public: Sol, Terra and Luna Explained
After a two-week government-requested delay, GPT-5.6 is now broadly available in three tiers — plus GPT-Live, a voice model that listens and speaks at the same time.
OpenAI Launches GPT-Live, a Full-Duplex Voice Model
GPT-Live-1 and GPT-Live-1 mini listen and speak at the same time — natural interruptions, live translation, and a smarter assistant working behind the scenes.
The White House Is Now Shaping Who Gets Access to Frontier AI Models
A new federal program, Gold Eagle, puts the U.S. government in the loop on which partners get early access to frontier models from labs like OpenAI and Anthropic.