Skip to main content
Back

Alibaba Releases 32B Reasoning Model to Rival DeepSeek R1

TRY NOW

March 6, 2025 6 min read

Hey,

Welcome to AlphaSignal – the most read newsletter by AI developers. 


We bring you the top 1% of news, papers, models, and repos, all summarized to keep you updated on the latest in AI.

IN TODAY'S SIGNAL

Read time: 6 min 02 sec

🎖️ Top News

📌  AI21

  • Deploy Jamba 1.6 now, the open-weight model, for faster, smarter, and secure AI with long-context and enterprise-ready performance.

⚡️ Trending Signals

💻  Top Lectures

🧠  Deep Dive

If you're enjoying AlphaSignal please forward this email to a colleague. 

It helps us keep this content free.

TOP NEWS

LLM

Qwen unveils an open-source 32B-parameter reasoning model, 4x faster with RL, rivals DeepSeek's 671B model

⇧ 10,835 Likes

What's New

Alibaba’s Qwen team releases QwQ-32B, a new reasoning model with 32 billion parameters, uses reinforcement learning (RL) to boost performance.

Despite its smaller size, it rivals models like DeepSeek-R1, which has 671 billion parameters. RL improves its abilities in math, coding, and general problem-solving tasks, providing a competitive edge in these areas.

Key Features

QwQ-32B utilizes RL to enhance task-specific performance.

  • Trains with outcome-based rewards for math, coding, and general reasoning tasks.

  • Includes accuracy verifiers and code execution servers for math and coding tasks.

  • Demonstrates continuous improvements during training across domains.

  • Maintains strong math and coding performance while scaling general capabilities.

Training

The model scales RL in two stages to refine its abilities.

  • First stage: RL trained for math and coding with accuracy verifiers and a code server.

  • Second stage: RL used for general capabilities like instruction following and agent performance.

  • Achieves continuous performance gains without significant trade-offs in math and coding.

Performance Benchmark

QwQ-32B scores high across key reasoning and coding benchmarks.

  • AIME24: 79.5%, matching DeepSeek-R1 in mathematical reasoning. A benchmark for assessing complex mathematical tasks.

  • LiveBench: 73.1%, surpassing distilled models in coding tasks. Evaluates coding and algorithmic performance through dynamic problem sets.

  • IFEval: 83.9%, ranking near DeepSeek-R1 in function evaluation. Tests models on functional programming and mathematical problem-solving.

  • BFCL: 66.4%, leading among similarly sized open-weight models. Focuses on logical reasoning and foundational cognitive tasks.

Availability and Access

You can use QwQ-32B with standard frameworks or cloud APIs.

  • Available on Hugging Face and ModelScope with Apache 2.0 licensing.

  • Supports Hugging Face Transformers for fine-tuning and local inference.

  • Requires high-memory GPUs due to its 32-billion parameter size.

  • Integrated with Alibaba Cloud DashScope API for hosted access.

Community Feedback

Farhan
"Smaller yet mightier, QwQ-32B - revolutionizing the AI world, one scaled down parameter at a time."

Dan Holzrichter

"I'm having troubles getting it to work correctly in vscode/cline... It seems to do a LOT of thinking, but not coding. I would love to have a local model for coding that is good."


Scott Hsieh
"After interacting with both QwQ-32B and DeepSeek-R1 using seven different questions as prompts, my impression is that QwQ-32B tends to get stuck in extended reasoning loops, often taking noticeably longer to complete its reasoning process."

Try now

Need an Open LLM that's Faster, Secure, and Optimized?

Jamba 1.6 is an open-weight AI model built for enterprise deployment. It outperforms Mistral, Llama, and Cohere in quality, speed, and long-context performance while ensuring data security.

  • Top Performance: Excels in RAG, long-context question answering, and overall model quality.

  • 256K Context Window: Processes long documents and conversations efficiently.

  • Hybrid Architecture: Mamba-Transformer MoE design improves cost and efficiency.

  • Flexible Deployment: Run on AI21’s SaaS or self-host for full data control.

  • Optimized for Speed: Jamba Mini 1.6 delivers low-latency responses for enterprise applications.

Deploy Jamba 1.6 in your environment today.

Try now

partner with us →

TRENDING SIGNALS

Coding Assistant

Codeium releases Windsurf Wave 4, adding AI-powered Previews for rapid app iteration and Claude 3.7 improvements

⇧ 3,047 Likes

Video Generation

Luma Labs introduces Keyframes, Extend, and Loop in Ray2 enhancing long-form video creation with better control

⇧ 4,869 Likes

LLM

OpenAI expands API access to o1 and o3-mini, adding vision, function calling, and Assistants API for all paid usage tiers

⇧ 1,159 Likes

Multimodal Models

Cohere presents Aya Vision 8B, an open-weights multimodal model that outperforms Llama 90B in image understanding

⇧ 1,048 Likes

Turing Award

ACM honors Barto and Sutton for their foundational work in RL, including temporal difference learning and policy-gradient methods

⇧ 1,468 Like

TOP LECTURES

Generative AI

How to build a ChatGPT-Powered AI tool to learn technical things fast

⇧ 451 Likes

Practical prompt engineering techniques to improve large language model (LLM) application performance. This lecture covers defining cognitive process boundaries, specifying input/output, implementing guardrails, and aligning workflows with human cognitive processes. Learn to leverage structured data, simplify processes, and iterate designs for optimized results in LLM-native apps.

LLM Evaluation

Instrumenting & Evaluating LLMs

⇧ 649 Likes

In this lecture, you'll learn how to instrument and evaluate LLMs. You'll explore unit tests for validating outputs, human evaluations to avoid overfitting, and tools like Langsmith for logging traces. You'll also cover common mistakes in LLM evaluation, metrics for agent performance, and strategies for improving AI assessments.

Agents

How to Get Your Data Ready for AI Agents (Docs, PDFs, Websites)

⇧ 1,846 Likes

Here you'll learn how to build an extraction pipeline for documents using parsing, chunking, and embedding techniques. You'll explore storing data in vector databases and searching it effectively. The video includes a practical demo, covering the creation of an interactive chat app to enhance AI responses through optimized data storage.

DEEP DIVE

Document Processing

Event-Driven Agentic Document Workflows

⇧ 1,247 Likes

This course teaches you how to automate document processing with an event-driven architecture. You’ll build an AI agent that fills out complex forms by retrieving information from source documents using Retrieval-Augmented Generation (RAG). The workflow runs asynchronously, processes multiple tasks in parallel, and integrates human feedback through text and voice.

You will learn to:

  • Design event-driven workflows with branching, looping logic, and concurrent execution.

  • Build an agent that parses documents, retrieves information using RAG, and structures responses.

  • Implement human-in-the-loop feedback for improved accuracy.

  • Set up a vector store and query engine for document retrieval.

  • Add multimodal capabilities to process spoken instructions.

By the end, you will have a functional, automated document processing workflow.

Start now

At Alpha Signal, our mission is to build a sharp, engaged community focused on AI, machine learning, and cutting-edge language models, helping over 200,000 developers stay informed and ahead. We’re passionate about curating the best in AI, from top research and trending technical blogs to expert insights and tailored job opportunities. We keep you connected to the breakthroughs and discussions that matter, so you can stay in the loop without endless searching. We also work closely with partners who value the future of AI, including employers and advertisers who want to reach an audience as passionate about AI as we are.


Our partnerships are based on shared values of ethics, responsibility, and a commitment to building a better world through technology.Privacy is a priority at Alpha Signal. Our Privacy Policy clearly explains how we collect, store, and use your personal and non-personal information. By using our website, you accept these terms, which you can review on our website. This policy applies across all Alpha Signal pages, outlining your rights and how to contact us if you want to adjust the use of your information. We’re based in the United States. By using our site, you agree to be governed by U.S. laws.

Looking to promote your company, product, service, or event to 200,000+ AI developers? Let's work together.

Work with us