Welcome back!

Alibaba’s QwQ-32B is proving that smart design can beat raw size. Despite being 20x smaller than DeepSeek-R1, it nearly matches its performance in math, coding, and reasoning, and runs smoothly on consumer GPUs. Let’s unpack...

Today’s Summary:

• 🚀 Alibaba’s QwQ-32B rivals DeepSeek-R1

• 📖 Mistral’s new OCR API

• 🤖 Amazon prepares Nova reasoning model

• ⚡ ChatGPT-4.5 now available for Plus users

• 🎬 OpenAI’s Sora launches in UK & EU

• 💻 ChatGPT for MacOS now edits code in IDEs

• 💰 OpenAI’s Premium AI Agents start at $2,000/month

• 👀 Google Gemini with vision launches in March

• 🛠️ 2 new tools

TOP STORY

Alibaba QwQ-32B rivals DeepSeek-R1, runs on consumer GPUs

The Summary: Alibaba Qwen has released QwQ-32B, a powerful reasoning AI model built using reinforcement learning (RL) techniques. Despite having 20x fewer parameters than DeepSeek-R1 (32B vs 671B), it nears the same performance on math, coding, and reasoning tasks.

Key details:

• Scored 78% on AIME 2024 math benchmark, beating Google Gemini 2.0 Flash

• Uses a 2-stage RL training prioritizing accuracy in math and coding tasks

• Processes inputs up to 131K tokens without performance degradation

• Requires only 24GB of vRAM compared to 1500GB for the full DeepSeek-R1

• Open-sourced on HuggingFace with Apache 2.0 license

Why it matters: As AI development faces diminishing returns from simply adding more parameters, QwQ-32B shows that smarter training methods can dramatically improve efficiency. If confirmed, this could push AI development toward more cost-effective and accessible models that run on local hardware.

Try QwQ-32B Demo

MISTRAL

New Mistral OCR API transforms document processing

The Summary: Mistral AI has released a new Optical Character Recognition (OCR) API that extracts content from images and PDFs with high accuracy. The system understands complex document elements including text, tables, images, and mathematical equations across multiple languages. Priced at 1,000 pages per dollar, Mistral OCR processes up to 2,000 pages per minute and outperforms competitors in benchmarks.

Key details:

• Processes up to 2,000 pages/minute on a single node

• Achieved 94.89% overall accuracy in benchmarks, beating Google, Azure, and OpenAI alternatives

• Transcribes thousands of fonts and languages with 99% accuracy

• Preserves document structure, including headers, paragraphs, tables, and mathematical formulas

• Can extract information as structured outputs like JSON

Why it matters: This advancement helps unlock the collective intelligence trapped in billions of unstructured documents. Organizations can transform their documents into searchable, analyzable assets that integrate with AI systems, turning static archives into dynamic intelligence networks.

AMAZON

Amazon to release Nova hybrid reasoning model

The Summary: Amazon is developing a new reasoning AI model with hybrid capabilities under its Nova brand, expected in June. The company has formed a dedicated group for agentic AI under AWS executive Swami Sivasubramanian, reporting to AWS CEO Matt Garman. The move comes as Amazon also prepares to release an updated Alexa with agentic capabilities later this month.

Key details:

• AWS CEO Matt Garman calls agentic AI "the next multi-billion business for AWS" in internal emails

• The new Alexa+ works in the background without user prompting, managing complex website interactions

• Amazon's upcoming Nova reasoning model aims to be 75% cheaper than competitors

• Amazon has already invested $8 billion in AI firm Anthropic while building its own competing models

Why it matters: This marks Amazon's move to claim territory in both consumer and business AI automation. Amazon's dual approach with Alexa+ for consumers and advanced reasoning models for AWS customers positions Amazon uniquely in the market for agentic systems that can complete tasks autonomously.

QUICK NEWS

Quick news

• ChatGPT-4.5 now available for all ChatGPT Plus users

• OpenAI’s Sora video generator is now available in the UK and EU

• ChatGPT for MacOS can now edit code directly in IDEs

• OpenAI plans Premium AI Agents starting at $2,000 monthly

• Google Gemini with vision to launch in March

TOOLS

🥇 New tools

• Octave TTS - Describe any AI voice and prompt its emotional delivery

• Lifestack - AI calendar using health data for better productivity