18 Jul 2026
Moonshot AI launched Kimi K3 as a frontier-class open-weights model with 2.8 trillion total parameters, 1M-token context, native multimodal input, and open weights promised by July 27, 2026. The model is live on Kimi.com, Kimi Work, Kimi Code, and API. According to Artificial Analysis, K3 ranks at 57 on the AA Intelligence Index, comparable to Opus 4.8 and GPT-5.5, and achieved #1 in Frontend Code Arena with 1679 points and a 76% pairwise win rate.
18 Jul 2026
Thinking Machines Lab introduced Inkling, a Mixture-of-Experts transformer with 975B total parameters and 41B active parameters that supports a 1M token context window. The model was pretrained on 45 trillion tokens of text, images, audio and video, and reasons natively across all three modalities. A smaller variant, Inkling-Small, with 12B active parameters was also previewed, with both models released as open-weights and available on Hugging Face, the company's Tinker platform, and partner services.
18 Jul 2026
PrismML released Bonsai 27B, based on Qwen 3.6 27B, in two compact variants under Apache 2.0 license: Ternary Bonsai 27B at 5.9 GB with 1.71 effective bits and 1-bit Bonsai 27B at 3.9 GB with 1.125 effective bits. The models preserve multimodal, tool-using, and long-context agentic workflows for local deployment, with demonstrations showing Hermes running on an RTX 5090 and phone deployment.
18 Jul 2026
Tencent Hunyuan released 1-bit and 4-bit quantized versions of Hy3, described as a 295B flagship-scale model that can be served on a single GPU via llama.cpp with MTP enabled. The release is part of a broader trend of aggressive compression bringing frontier-adjacent models onto consumer devices and enabling local inference for agentic workflows.
18 Jul 2026
Ant Group's AI Safety Lab has open-sourced SingGuard-NSFA, a safety guardrail model designed to detect risks in autonomous agents including prompt injection, sensitive data theft, malicious code execution, resource abuse, and permission misuse. The model covers seven major risk categories, 28 subcategories, 185 scenarios, and 133 languages, with nearly 100,000 samples in its evaluation system. SingGuard-NSFA is available in 0.8B, 2B, 4B, and 9B versions and can make a single risk judgment in about 50 milliseconds.
18 Jul 2026
Meta released Muse Spark 1.1, combining a million-token context window with multimodal perception, coding, computer use, and multi-agent orchestration. The model features active context management that compacts extended sessions while preserving necessary state. Meta also launched a paid Model API, marking a strategic shift toward selling metered intelligence.
18 Jul 2026
OpenAI released GPT-5.6 in three model sizes named Sol, Terra, and Luna, along with a new ultra effort level that coordinates multiple agents in parallel for complex tasks. Terra performs at GPT-5.5-like capability at lower cost while Luna offers the fastest and cheapest option, with both achieving higher performance than competing models at roughly one-third the time and one-quarter the estimated cost. The launch also included updates to ChatGPT Work, a new desktop app merging Codex and ChatGPT, along with new features like Sites beta and multi-agent capabilities in the Responses API.
18 Jul 2026
Anthropic has created a tool called the Jacobian lens (J-lens) that uncovers a hidden area inside Claude Opus 4.6 named J-space, which contains words related to what the model is likely to produce in the near future. The technique builds on previous mechanistic interpretability research and reveals that what language models actually do can differ from what they claim to be doing. Anthropic shared its results in a paper this week and partnered with Neuronpedia to provide a public demo.
18 Jul 2026
xAI publicly launched Grok 4.5, positioning it as an Opus-class model that is faster, more token-efficient, and lower cost than existing alternatives. The model was trained specifically for coding and agents in partnership with Cursor, which announced in-product availability with double usage for the first week. Grok 4.5 became available across Grok Build, API, Cursor, and other platforms including OpenRouter and Hermes Agent on day one.
18 Jul 2026
Alibaba released HappyHorse 1.1, an updated video generation model, and launched an AI film competition with Huajing Entertainment Group. The company said the new version improves on version 1.0 across several areas.
18 Jul 2026
Alibaba's Qwen team released a robotics suite of three foundation models, Qwen-RobotNav, Qwen-RobotManip, and Qwen-RobotWorld. Each aligns language with a different class of physical action, extending vision-language models into navigation and manipulation.
