Kimi K3, Inkling, and GPT-Red: July 2026 AI Releases

Three significant AI model releases landed in July 2026: Moonshot AI published a 2.8-trillion-parameter open-source model, Thinking Machines debuted its first multimodal open-weights system from former OpenAI CTO Mira Murati, and OpenAI detailed an automated red-teaming model used to harden GPT-5.6 against prompt injection attacks.

Moonshot AI’s Kimi K3 Claims Open-Source Record

Beijing-based Moonshot AI released Kimi K3 on Thursday, describing it as the largest open-source AI model ever built at 2.8 trillion parameters — roughly 75% larger than DeepSeek’s V4 Pro, which Moonshot’s own technical documentation places at approximately 1.6 trillion parameters. According to Moonshot’s quickstart documentation, the model includes a 1-million-token context window and native visual understanding. Full model weights are scheduled for release on July 27, per details shared with researchers who reviewed the company’s technical materials.

Moonshot, which is backed by Alibaba, timed the launch ahead of the 2026 World Artificial Intelligence Conference in Shanghai. The release marks a notable recovery for the company, whose market position had eroded over the prior 18 months following DeepSeek’s rise, according to VentureBeat’s coverage. Benchmark results cited by Moonshot show Kimi K3 performing close to proprietary systems from Anthropic and OpenAI, though independent third-party verification of those claims was not available at publication.

The model is currently accessible at kimi.com with a Google account or phone number — no payment required.

Thinking Machines Releases Inkling Under Apache 2.0

Thinking Machines — the AI startup founded by former OpenAI CTO Mira Murati — released its first major model, Inkling, on an Apache 2.0 open-source license, targeting enterprises that want to run capable models on-premises or in private cloud environments. According to Thinking Machines’ announcement, Inkling is a natively multimodal Mixture-of-Experts system with 975 billion total parameters, capable of reasoning across text, images, and audio.

On third-party benchmarks, Inkling scored 77.6% on SWE-bench Verified — above Nvidia Nemotron 3’s 71.9% — and 91.4% on VoiceBench, compared to 94.4% for Gemini 3.1 Pro at high reasoning effort. The company says the model was designed with a “controllable thinking effort” mechanism to balance cost against performance, a departure from fixed-compute inference strategies used by some frontier competitors.

Thinking Machines also flagged a design choice unusual among enterprise models: Inkling was built “to answer directly on topics that may be subject to censorship,