Qwen3 Trails Closed Model Rivals

PLUS: Amazon Launches Nova Premier Model, Microsoft To Host Grok AI and more.

Author

Wes Roth
May 02, 2025

In partnership with

There’s a reason 400,000 professionals read this daily.

Join 400,000+ professionals at Google, Microsoft, OpenAI, and more who start their day with The AI Report — the #1 B2B AI newsletter. We cut through the noise and show you exactly how to use AI tools to save time, boost productivity, and stay ahead at work.

No hype. No fluff. Just real-world use cases, expert prompts, and actionable tools.

Today:

  • Qwen3 Trails Closed Model Rivals

  • Microsoft Releases Phi-4 Reasoning Model

  • Claude Gains Integrations And Research Mode

  • Amazon Launches Nova Premier Model

  • Microsoft To Host Grok AI

Is Qwen3 the new CODING KING? (model testing)

Qwen 3 (235B A22B) impressed with a detailed HTML solar system simulation and a working reinforcement learning pipeline for Snake. 

While it handled some prompts well, it often required extra tweaking and fell short of Gemini 2.5 Pro, Claude 3.7, and OpenAI o3 in reliability and polish. It’s strong—likely the best open-source coding model—but not yet a top-tier rival to closed models from Google, OpenAI, or Anthropic.

Microsoft released Phi-4-Reasoning-Plus, a small but powerful open-weight AI model focused on deep reasoning. It uses supervised fine-tuning and reinforcement learning to outperform much larger models on math, logic, and coding tasks. Trained on carefully curated data, it emphasizes transparency and step-by-step thinking. With a 14B parameter size and MIT license, it’s optimized for both performance and accessibility across research, enterprise, and safety-sensitive environments.

Why This Matters

  1. High Reasoning in Small Models – It proves that smaller models, when trained well, can rival or beat much larger ones in complex reasoning tasks.

  2. Open and Flexible Deployment – Its permissive license and compatibility with major frameworks enable wide use in commercial and research contexts.

  3. Safe and Interpretable Outputs – Designed for transparency and alignment, it supports responsible AI deployment in regulated or high-stakes settings.

Anthropic launched Claude Integrations, letting users connect Claude to tools like Jira, Zapier, and Intercom, giving it full context of work apps and allowing it to take actions across platforms. Claude also gained a powerful Research mode, which performs deep investigations across web, Google Workspace, and connected tools, delivering reports with sources in under 45 minutes. These updates make Claude a smarter, action-ready assistant for complex tasks and enterprise workflows.

Why This Matters

  1. Turns Claude into an AI Agent – With tool integration and task execution, Claude shifts from chatbot to capable AI assistant.

  2. Expands LLM Utility in Enterprises – Claude now supports deep workflow automation, research, and team collaboration across real business apps.

  3. Sets a New Benchmark for AI Context Awareness – Claude can now understand full project histories and data from multiple platforms, improving response relevance and reliability.

Amazon has launched Nova Premier, its most advanced AI model yet, designed for complex reasoning, multimodal input, and large-scale workflows. With a 1 million token context window, it excels in tasks like investment research using multi-agent systems. Nova Premier can also act as a teacher model to create smaller, faster versions like Nova Micro via model distillation, making advanced AI capabilities more cost-effective and deployable across production environments using Amazon Bedrock.

Why This Matters

  1. Unmatched Context Handling – With support for 1 million tokens, Nova Premier pushes the boundary for long-context understanding in enterprise AI applications.

  2. Model Distillation at Scale – It enables practical, production-grade AI by transferring top-tier capabilities into smaller, optimized models for real use cases.

  3. Multi-Agent AI Orchestration – Nova Premier powers sophisticated agent collaboration, marking a shift from single-model outputs to AI systems managing entire workflows.

🧠RESEARCH

Sadeed is a lightweight Arabic language model that significantly improves text diacritization—a tough challenge due to the language’s complexity. Fine-tuned on clean, high-quality data, it rivals larger models while using fewer resources. The paper also introduces a new benchmark, SadeedDiac-25, to better evaluate diacritization across different text types.

Phi-4-reasoning is a 14B-parameter model built for complex thinking tasks. Fine-tuned on carefully chosen prompts and boosted with reinforcement learning, it generates clear step-by-step reasoning. It beats much larger open models and rivals top-tier systems in math, coding, planning, and science tasks—proving that smart training matters more than just model size.

Foundation-Sec-8B is a cybersecurity-focused language model built on Llama 3.1 and trained with specialized security data. Despite its smaller size, it matches much larger models like Llama 3.1-70B and GPT-4o-mini on key cybersecurity tasks. The open release aims to push forward AI adoption in public and private security sectors.

RoboVerse is a unified platform for robot learning that combines simulation tools, a large synthetic dataset, and standardized benchmarks. It supports various simulators and robot types through a universal interface called MetaSim. By improving data quality and evaluation consistency, RoboVerse helps boost performance in imitation learning, reinforcement learning, and sim-to-real transfer.

🛠️TOP TOOLS

Facetune - AI-powered tools for enhancing, reshaping, and creatively transforming digital content, empowering users to express their unique style and vision in every post.

TextCortex AI - AI-powered writing assistant designed to streamline content creation across various platforms and formats.

Neverinstall - Cloud-based platform that provides instant access to high-performance virtual desktops and cloud PCs directly through a web browser.

Tome - AI-powered storytelling and presentation platform that has evolved to focus primarily on sales and marketing professionals.

Wav2Lip - All-in-one extension designed to generate high-quality lip-sync videos by enhancing the output of the original Wav2Lip tool.

📲SOCIAL MEDIA

🗞️MORE NEWS

  • Microsoft is preparing to host Elon Musk’s Grok AI on Azure, expanding its support for rival models beyond OpenAI. The move could strain its OpenAI partnership and signals Microsoft’s push to dominate AI infrastructure.

  • Ai2 released Olmo 2 1B, a small open-source AI model that outperforms similarly-sized models from Google, Meta, and Alibaba on reasoning and truthfulness tests. It runs on basic hardware and is fully reproducible.

  • Midjourney has upgraded its image model v7 with better image quality, improved hand and body rendering, and smarter prompt alignment. It also added new editing tools, a redesigned interface, and an --exp parameter for more control over visual detail.

  • Google’s Gemini app now includes built-in image editing tools, letting users modify or generate images by changing backgrounds, objects, or appearances. The feature is rolling out gradually in over 45 languages and most countries.

  • Meta’s Ray-Ban smart glasses now record voice by default to train its AI. Users can’t fully opt out, only delete individual recordings or disable voice control. Camera footage isn’t used for AI unless shared. Privacy concerns are growing.

  • Amazon Web Services is building a new AI coding tool to rival startups like Cursor and Windsurf. It will combine editing, testing, and debugging in one app, aiming to boost developer productivity and drive more AWS cloud usage. 

  • A new study claims AI benchmark LMArena unfairly benefits big players like OpenAI and Meta by letting them test and submit only top-performing model versions. Researchers also flagged unequal data access and hidden model removals. LMArena denies bias but may review its submission and transparency policies.

  • Google’s NotebookLM app is now available for pre-registration on both the App Store and Google Play, with a release date expected around May 20th.

What'd you think of today's edition?

Login or Subscribe to participate in polls.

Reply

or to participate.