Summary
Wes Roth covers a dense 48-hour window of AI model releases and competitive developments. The headline story is Kimi K3 from Moonshot AI, which appears to have soft-launched following a promo page briefly going live with a July 15th date. Leaked specs describe a 2.5-trillion-parameter model with a 1-million-token context window and a novel architecture whose CEO claims targets capabilities other models haven’t defined. Early arena comparisons suggest Kimi K3 may be competitive with Grok/Fable 5 on visual coding tasks — which Roth frames as a potential second DeepSeek moment for the industry.
The episode also covers a trillion-parameter model released by Meera Miai, a former OpenAI researcher; OpenAI’s GPT-RED, a model purpose-built for adversarially attacking other AI systems; and rumors of challenges surrounding Google’s Gemini 3.5 Pro. A significant portion analyzes Microsoft’s MAI model family alongside a separate open-source model called Inkling, both positioned for enterprise fine-tuning — where organizations use reinforcement learning on proprietary data to build domain-specific models that stay in-house, appealing particularly to regulated industries like healthcare that cannot share internal data with frontier labs.
Roth frames the broader competitive picture as a divergence between two strategies: racing to maintain generalist frontier leadership versus offering a fine-tunable foundation that corporations can make their own — a tradeoff with major implications for enterprise AI adoption and data privacy.
📺 Source: Wes Roth · Published July 16, 2026
🏷️ Format: News Analysis
