🚨 AI News | TestingCatalog (@testingcatalog) on X

What it claims

This is a vendor-grouped news digest, composed with Grok, covering Oct 17 and part of Oct 16. It does not establish any result of its own. Its value over the summary lies in the access terms, sizes and context lengths it attaches to each item:

• Claude Code Review (GA): available only on Team and Enterprise plans. It has three review levels (Quick, Standard and Deep), which are configured per repository.
• GPT-6 Astra Pro: the upgrade is limited to ChatGPT Pro users. It adds faster responses, better plugin and tool use, and multiple agents working in parallel.
• Codex and Work: "more effective usage" comes from three changes. Context is reused across subagents, cached for longer, and requests are routed to get more cache hits.
• Gemini 4 Argon: in AI Studio Build and the playground for Google AI Pro subscribers. It is also in Perplexity for Pro and Max subscribers. AI Studio Build added two modes: Plan, which plans before making changes, and Max, which works longer for better results.
• NVIDIA Nemotron Labs Dory 1.0: a 4B looped open model with a 256K context window. It was released quietly on Hugging Face.
• Perplexity Decider v1.2 / Decider Flash v1.2: these are multimodal "decision" models at 27B and 4B.
• DeepSeek-V4.1-Flash-1017: MIT-licensed. Besides the claimed gains, it promises faster, shorter outputs.
• Reflection Beam: Apache 2.0, with 40+ launch partners.
• Tencent Hy Memory 2.0: ships as a memory model, a framework and the MemHub app. MemHub syncs across 10+ AI tools.
• Meta Muse Spark Open: distributed on Hugging Face. It is described as optimized to run locally on a Mac Studio, despite being 816B total / 42B active.
• Grok Bots: now on web and mobile. Computer tasks finish 3x sooner on average, and other replies take less than half the time.
• NVIDIA RTX Spark Windows laptops: a Blackwell GPU paired with a Grace CPU and up to 128GB of unified memory. They are now on sale.
• Cursor cloud agents: can build interactive canvases such as dashboards and reports, and share them as a link.

How to use it

Use this as a pointer list, not a source. The items that can be acted on directly:

• Small long-context open model: look for Dory 1.0 on Hugging Face.
• Permissive licences: DeepSeek-V4.1-Flash-1017 is MIT and Beam is Apache 2.0.
• Local frontier-scale MoE: Muse Spark Open, with Mac Studio as the stated target.
• Plan gating to check first: Claude Code Review requires Team or Enterprise, and Astra Pro requires ChatGPT Pro.

The post gives no repository names, model IDs, commands or links.

Limits

• No baselines: the post itself says Grok composed it, with human cherry-picking and post-editing. Every performance claim lacks a baseline or setting, including "big gains", "3x sooner" and "faster responses". None of these count as evidence.
• Unconfirmed item: the Gemini Notebook Slides/Video item rests on a hidden notice and is described as "preparing", so it is not a release.
• Truncated page: the text cuts off inside the Oct 16 brief, mid-sentence on "Nemotron 3.5 Super VL, an open…". The rest of Oct 16 is missing.
• Licences unstated: the post gives no licence for Dory, Decider, Muse Spark Open or Hy Memory. "Open" and "open-source" are its only words for them, so check each model card before relying on them.