Roundup of Recent Frontier Model Releases

johnseach · x · 2026-07-19

This post summarizes the most important batch of model releases from early to mid-July 2026. The core theme is that **Chinese and US frontier models continue to advance, while competition over open weights and pricing intensifies**. ### Key Releases - **Kimi K3 (Moonshot)**: Approx. **2.8T parameters**, multimodal, **1M context**. Overall performance is close to Claude Fable 5 and GPT-5.6 Sol; particularly strong in frontend/Web development and certain coding benchmarks. Weights are planned to be open-sourced on **July 27**, with aggressive pricing. - **Inkling (Thinking Machines Lab / Mira Murati)**: Approx. **975B total parameters**, MoE architecture, **41B active parameters**. Supports text/image/audio and used video data during training; features a **1M context** and **Apache 2.0 open weights**. Positioned more as a differentiable, customizable foundation model. - **Qwen 3.8 / Qwen3.8-Max-Preview (Alibaba)**: Approx. **2.4T parameters**. Initially accessible via products like Token Plan, with full open weights expected soon. Officially claimed to perform just behind Claude Fable 5, approaching or exceeding recent frontier models. - **GPT-5.6 Series (OpenAI)**: Rolled out in batches starting July 8–9, with continued expansion of availability. - **Grok 4.5 (xAI)**: Released on July 8. - **Muse Spark 1.1 (Meta)**: Released on July 9, offering a public Meta Model API preview, focusing on stronger agentic/coding capabilities and lower prices. ### Overall Assessment The author believes that the main narrative of this week is: **Chinese models are closing the gap and have even taken the lead in certain niches, while open/near-open weights have become a crucial competitive frontier**. On the other hand, price pressures and the debate between open and closed-source routes remain central industry issues.

Original post →

More from Models

Models channel →