Open Weights AI: The Escalation Edition
A 2.8-trillion-parameter model with a promised weights release, a call for openness from Beijing and a new open-weights release from Thinking Machines Lab are redrawing the frontier. One analyst sees a narrower performance gap if the promised release occurs, while the geopolitics are heating up and the definition of “open” remains demanding.
In brief
Moonshot AI says its 2.8-trillion-parameter Mixture-of-Experts model Kimi K3 will receive a weights release by July 27, 2026 . Alibaba has reversed course by releasing Qwen 3.8 Max as open weights . Thinking Machines Lab released Inkling, an Apache-2.0-licensed model with 975 billion total parameters . Policy debates are intensifying around distillation and unconfirmed reports of White House discussions about managing open models through an executive order .
Kimi K3: A Colossus Awaiting Its Weights
Moonshot AI announced Kimi K3 on July 16 as a 2.8-trillion-parameter Mixture-of-Experts model, with its weights promised by July 27, 2026 . Moonshot calls it the first “open 3T-class model”, rounding 2.8 trillion parameters to the 3-trillion class; by total parameter count, it is larger than DeepSeek’s 1.6-trillion-parameter v4 Pro .
On Artificial Analysis’s private long-horizon knowledge-work evaluation, K3 achieved an Elo of 1547 and ranked behind only Claude Fable 5 . K3 also used 21% fewer output tokens than K2.6 on the Artificial Analysis Intelligence Index .
Assuming Moonshot fulfils its promised weights release, Interconnects AI estimates that either the open-to-closed or American-to-Chinese model-performance gap has narrowed from a debated six to nine months to roughly three to five months .
Beijing’s Call for Openness
At the World AI Conference, President Xi Jinping encouraged openness and open source . He urged the industry to “seize this rare, historic opportunity to encourage open source, openness, collaboration and sharing” .
Alibaba’s release of Qwen 3.8 Max as open weights reversed its May decision not to release Qwen 3.7 Max that way .
Inkling: A Release with Caveats
Thinking Machines Lab released Inkling, an Apache-2.0-licensed multimodal Mixture-of-Experts model with 975 billion total parameters and 41 billion active parameters .
Inkling is intended as a strong base model for fine-tuning through the lab’s Tinker training platform . Simon Willison described its model card and training-data documentation as unusually sparse and offering little insight into the data sources .
The Distillation Battle and Looming Regulation
Interconnects AI says multiple sources are citing White House discussions about a new executive order to manage open models, although no official information has been released; the report says any measure would likely affect Chinese-origin models and government uses .
Ben Thompson has proposed that the US explicitly make data collection for model training fair use and bar terms of service that forbid distillation, at least for US companies . Interconnects AI argues that the anti-distillation campaign amounts largely to regulatory capture because the proposed solutions would benefit the organizations advocating them .
Ecosystem in Motion: TTS and the Definition of Open
Mistral released Voxtral TTS, which it describes as a frontier open-weights text-to-speech model for voice agents, and also introduced Mistral Small 4 . Hugging Face and Cerebras are bringing Gemma 4 to real-time voice AI .
The Open Source Initiative’s Open Source AI Definition 1.0 says that “Open Source models” and “Open Source weights” must include the data information and code used to derive their parameters .
Sources
- Open models recap: more on Kimi K3, Qwen 3.8, Xi's WAIC speech, distillation, the open-closed gap, and what's next — Interconnects AI, 2026-07-22
- The Open Source AI Definition – 1.0 – Open Source Initiative — opensource.org
- Latest news | Mistral — mistral.ai
- b10091 — ggml-org/llama.cpp, 2026-07-22
- b10090 — ggml-org/llama.cpp, 2026-07-22
- b10089 — ggml-org/llama.cpp, 2026-07-22
- Hugging Face and Cerebras bring Gemma 4 to real-time voice AI — Hugging Face - Blog, 2026-07-01
- Who’s Afraid of Chinese Models? — Simon Willison's Weblog, 2026-07-20
- Kimi K3: The open-weights escalation — Interconnects AI, 2026-07-20
- Kimi K3, and what we can still learn from the pelican benchmark — Simon Willison's Weblog, 2026-07-16
- Inkling: Our open-weights model — Simon Willison's Weblog, 2026-07-16
- 6 months to live for open models — Interconnects AI, 2026-07-12