Moonshot’s Kimi K3 Becomes the Largest Open Model, Thinking Machines Debuts Inkling, and OpenAI Details GPT-Red Adversarial Safety Training

Posted on by

This brief covers the trailing ~72 hours (July 15–18, 2026). Every item below was confirmed on the originating organization’s own page, with a published date inside the window. It was a heavyweight stretch for open-weight models: Moonshot shipped the largest open model ever announced, and Mira Murati’s Thinking Machines Lab released its first model, while OpenAI published three posts spanning safety research, teen policy, and AI economics, and Google DeepMind and Isomorphic Labs laid out a joint biosecurity strategy.

Moonshot AI launches Kimi K3, a 2.8-trillion-parameter model it calls the first open 3T-class model

Moonshot AI · July 16, 2026

Moonshot introduced Kimi K3, a 2.8T-parameter Mixture-of-Experts model (16 of 896 experts active) built on its Kimi Delta Attention and Attention Residuals architectures, with native vision and a 1-million-token context window. Moonshot says K3 trails only Claude Fable 5 and GPT-5.6 Sol overall while consistently outperforming other tested models, and it is priced at $3/$15 per million tokens — the most expensive Chinese-lab model to date. K3 is live on Kimi.com, Kimi Work, Kimi Code, and the Kimi API, with full open weights promised by July 27, 2026.

“Today, we are introducing Kimi K3 — our most capable model. Kimi K3 is a 2.8T-parameter model built on our Kimi Delta Attention and Attention Residuals, with native vision capabilities and a 1-million-token context window. It is the world’s first open 3T-class model.” — Moonshot AI

Source: Kimi K3: Open Frontier Intelligence

Thinking Machines Lab releases Inkling, its first open-weights model

Thinking Machines Lab · July 15, 2026

Mira Murati’s Thinking Machines Lab released Inkling, a 975B-parameter Mixture-of-Experts model (41B active) trained from scratch on 45 trillion tokens of text, images, audio, and video, with a 1M-token context window and controllable thinking effort. The lab positions Inkling not as a benchmark leader but as a broad, balanced base for customization via its Tinker fine-tuning platform, and it trained the model specifically for calibration, instruction following, and resistance to censorship. Full weights are on Hugging Face, and a lighter Inkling-Small (12B active) is previewed.

“Inkling is not the strongest overall model available today, open or closed. Instead, a combination of qualities makes it a good open-weights base for customization: multimodal capabilities, efficient thinking, and availability on Tinker for fine-tuning.” — Thinking Machines Lab

Source: Inkling: Our open-weights model

OpenAI reveals GPT-Red, an internal red-teaming model trained via self-play to harden GPT-5.6

OpenAI · July 15, 2026

OpenAI detailed GPT-Red, an internal-only automated red-teaming model trained with self-play reinforcement learning at the compute scale of some of its largest post-training runs. GPT-Red found successful prompt-injection attacks in 84% of held-out scenarios versus 13% for human red-teamers, and even broke a live Andon Labs vending-machine agent in OpenAI’s office. Used adversarially during GPT-5.6’s training, it helped drive the model’s failure rate on GPT-Red’s direct prompt injections down to 0.05%, with a pre-print promised the following week.

“We believe automated red-teaming unlocks a crucial form of self-improvement for safety: using today’s models to directly help make future models safer.” — OpenAI

Source: GPT-Red: Unlocking Self-Improvement for Robustness

Google DeepMind and Isomorphic Labs publish a joint approach to bioresilience

Google DeepMind · July 16, 2026

Google DeepMind and Isomorphic Labs published a joint bioresilience framework organized around prevention, detection, and response: adapting SynthID watermarking to biological sequences for DNA-synthesis screening, using AlphaEvolve to cut the cost of metagenomic pathogen surveillance, and granting trusted researchers access to frontier systems to accelerate vaccine and countermeasure design. Isomorphic Labs has stood up a dedicated unit to rapidly deploy its Drug Design Engine during novel outbreaks, and the pair report more than 15 partnerships with governments and biosecurity organizations over the past 12 months.

“Our work is twofold – to prevent threat actors from misusing our models, and to ensure that governments, scientists, biosecurity experts and our teams can harness these technologies to build a more resilient world.” — Google DeepMind and Isomorphic Labs

Source: Our approach to bioresilience

OpenAI argues teens deserve access to safe AI, expands Study Mode parental controls

OpenAI · July 16, 2026

OpenAI published its case for teen access to AI paired with age-appropriate protections, citing that nearly 9 in 10 teens on ChatGPT use it weekly for learning or productivity. New measures include letting parents enable Study Mode by default from Parental Controls, education-focused starter prompts, more frequent break reminders for teens, and expanded parent notifications — now covering account deactivations for violent-threat policy violations, developed with violence-prevention firm Moonshot (unrelated to Moonshot AI). OpenAI also announced it has joined the Family Online Safety Institute.

“Keeping teens from using it until adulthood would be like asking a previous generation to avoid the internet or search engines until they turned 18, leaving them less prepared to use one of the defining technologies of their time.” — OpenAI

Source: Why teens deserve access to safe AI

OpenAI CFO Sarah Friar proposes a four-part “scorecard for the AI age”

OpenAI · July 17, 2026

In a follow-up to last week’s enterprise spend playbook, OpenAI CFO Sarah Friar proposed measuring AI ROI as “Useful Intelligence per Dollar” across four questions: how much useful work gets done, what a successful task costs, how dependable the results are, and whether each AI dollar buys more work at scale. The post frames GPT-5.6’s three tiers (Sol, Terra, Luna) as levers in that equation and claims GPT-5.6 Sol set a new state of the art on the Artificial Analysis Coding Agent Index while using 54% fewer output tokens than another leading model.

“The ultimate scorecard for the age of AI could be looked at as ‘Useful Intelligence per Dollar.’” — Sarah Friar, CFO, OpenAI

Source: A scorecard for the AI age

This brief covers the trailing ~72 hours (July 15–18, 2026).

Primary sources:

Leave a Reply