Kimi K3 and Inkling: Open Weights Hit 3 Trillion Parameters
July 2026 was the month open weights stopped being the budget option and started being the scale option. Moonshot shipped Kimi K3, the first open 3-trillion-class model, at 2.8 trillion parameters. Thinking Machines followed with Inkling, a 975-billion-parameter Apache 2.0 model that now holds the title of top US open-weights release. Between them, they changed what teams can expect to run without a frontier vendor.
Kimi K3: the first open 3T-class model
Kimi K3 is a mixture-of-experts model with 2.8 trillion total parameters, activating roughly 104 billion per token. It pairs a new attention scheme, Kimi Delta Attention, with attention residuals for roughly 2.5x the scaling efficiency of K2, adds native vision, and ships a 1M-token context window. Pricing landed around half of Opus 4.8 or GPT-5.6 Sol per token. At release it debuted number one on the Frontend Code Arena, above Claude Fable 5, and third on Artificial Analysis' Intelligence Index. Demand was high enough that Moonshot paused new API subscriptions, and the full weights shipped on July 27 under a bespoke open-weight license, not an OSI-approved one.
Inkling: the top US open-weights model
Mira Murati's Thinking Machines released Inkling under Apache 2.0: a 975-billion-parameter, 41-billion-active mixture-of-experts transformer, pretrained from scratch on 45 trillion tokens of text, images, audio, and video. It carries a 1M-token context window, uses an encoder-free design with native reasoning across all three modalities, and ranks 41st on the Artificial Analysis Intelligence Index, the best US open-weights result so far. A leaner Inkling-Small variant at 276 billion total parameters, with 12 billion active, was previewed alongside it, and both run on the Tinker platform at a limited-time 50% discount.
What the open-weights surge changes
The two releases are part of a broader shift: Chinese open-weight models now account for roughly 30% of global usage, up from 1.2% eleven months ago, and open-model cloud platforms are scaling with them. Together AI closed an $800M Series C at an $8.3B valuation on the back of more than $1B in annual bookings, with open-model usage on its platform tripling year over year. Meituan's LongCat-2.0, a 1.6-trillion-parameter model trained entirely on Chinese ASICs without NVIDIA hardware, runs at $0.038 per million tokens with free cache hits, a price point that would have been unthinkable for a model of its size a year ago. The pattern is clear: the cost floor for serious AI work is being set by open models, and frontier vendors are responding with aggressive price cuts of their own.
The broader open-weights surge
Kimi K3 and Inkling are the largest releases, but they sit inside a wave. Chinese open-weight models now account for roughly 30% of global usage, up from 1.2% just eleven months ago. Meituan disclosed LongCat-2.0, a 1.6-trillion-parameter MoE trained entirely on Chinese ASICs with no NVIDIA hardware, scoring 59.5 on SWE-bench Pro at $0.038 per million tokens with free cache hits. Alibaba previewed Qwen3.8-Max at a claimed 2.4 trillion parameters, with Alibaba shares rising as much as 5.4% on the announcement day. Together AI closed an $800 million Series C at an $8.3 billion valuation, with open-model usage on its platform tripling year over year and annual bookings exceeding $1 billion. The pattern is not just that open weights are competitive; it is that the infrastructure layer funding them is scaling faster than the model layer itself.
The API pause and what it signals
Moonshot was forced to pause new Kimi K3 API subscriptions within days of launch because demand exceeded provisioned capacity. That is unusual for an open-weights model, where the expectation is that users download weights and self-host rather than hitting a rate-limited API. The reality is that running a 3T-class model, even one with only ~104B active parameters, requires inference infrastructure that most teams do not maintain in-house. The API pause means the model's availability for real-world use still trails its benchmark table position, and teams evaluating Kimi K3 should check whether API access has resumed or whether self-hosting is feasible on their infrastructure before including it in production plans.
What to watch
The licensing question matters as much as the parameter count. Kimi K3's bespoke license is not OSI-approved, which limits some enterprise uses, while Inkling's Apache 2.0 is broadly permissive. Both run fine in the API tier for most teams, but self-hosting decisions should be made against the license text, not the benchmark table. On capability, the caveat is effort: open models at this scale match frontier closed models on coding and reasoning benchmarks, but system-level safety tooling, serving reliability, and support remain areas where vendors still earn their margin.
For teams evaluating open weights, the practical test is your own workload: put Kimi K3 or Inkling behind your eval set, compare total cost including hosting and engineering time, and only then compare headline benchmark numbers. The era where open weights meant a meaningful quality sacrifice is over; the era where they are the default production choice has not quite arrived. For teams that cannot self-host, Together AI and other open-model clouds now serve both Kimi K3 and Inkling through standard APIs, making the open-weights advantage accessible without managing inference infrastructure. As the open-weights ecosystem matures, the distinction between frontier closed and open models will be about licensing and control, not capability. The key takeaway for teams is that open weights have crossed the scale threshold where self-hosting a frontier-class model is not just viable but cost-effective, and the remaining gap with closed vendors is not model quality but the operational burden of inference infrastructure, safety tooling, and ongoing maintenance.