Skip to main content
Back

Alibaba Qwen3-Next beats 32B at <10% training cost

Claude Team gets Memory, Cursor's Tab Upgrade, Hugging Face LLMs Inference, OpenAI nonprofit gains $100B stake.

September 12, 2025 4 min read

Hey

Welcome to AlphaSignal, the most read source of news by AI engineers and researchers.


Every day, we identify and summarize the top 1% of news, papers, models, and repos, so you’re always up to date.


Here’s today’s roundup:

SUMMARY

Read time: 4 min 25 sec

Top News

▪ Alibaba releases a 80B hybrid model with faster, cheaper long-context training

Encord

▪ Exclusive Access: Physical AI Hub for ML Leaders

Trending Signals

▪ Anthropic launches memory in Claude for Team and Enterprise

▪ Cursor upgrades its inline code suggestion system

▪ Hugging Face lets VS Code users run open LLMs in Copilot Chat

▪ OpenAI, Microsoft sign MoU; nonprofit gains $100B+ equity stake

▪ VS Code improves coding agent integration in the new update

Top Reads

▪ Anthropic shows how to prototype and refine agent tools

▪ Google DeepMind explains GPU memory, networking, and collectives

▪ Hugging Face brings GPT-OSS tricks to Transformers

Top Tutorial

▪ Google tutorial shows how to build MCP servers with Gemini CLI

TOP_NEWS

Alibaba's Qwen debuts Qwen3-Next, delivering 10× faster long-context inference with hybrid attention design

4,275 Likes

alpha_signal_image_2

Alibaba announced Qwen3-Next, a new architecture designed to push efficiency in both long-context training and inference. Unlike the MoE-based Qwen3, the new design introduces a hybrid attention system, ultra-sparse MoE routing, stability-focused optimizations, and native multi-token prediction.


Key Features

  • Handles very long inputs (up to 256K tokens, tested to 1M).

  • Uses multi-token prediction to speed up text generation.

  • Needs far less compute than Qwen3-32B for training.

  • Runs 10× faster than Qwen3-32B on long texts.

Technical Architecture

  • Hybrid attention: Mixes fast linear attention with standard attention for better balance of speed and recall.

  • Ultra-sparse MoE: 80B parameters total, but only ~3B active at a time, making it efficient.

  • Stability tweaks: Uses improved normalization and gating methods to avoid training crashes or unstable weights.

  • Speculative decoding: Built-in multi-token prediction improves generation speed and consistency.

Performance

  • The base model (80B total, 3B active) matches or beats Qwen3-32B while using under 10% of its training cost.

  • Inference speed is up to 10× faster for context lengths over 32K tokens.

  • The instruct version performs close to Qwen3-235B and handles ultra-long context (up to 256K tokens) better.

  • The thinking version outperforms Qwen3-30B/32B and Gemini 2.5 Flash Thinking, nearing Qwen3-235B performance.

Access & Availability

  • Released on Hugging Face and ModelScope.

  • Available in Instruct and Thinking variants.

  • Supported by vLLM and SGLang for OpenAI-compatible serving.

  • Integrated with Qwen-Agent for tool calling and agentic workflows.

Try now
alpha_signal_image_3

Get Exclusive Access to Physical AI and LiDAR Resources 

Visit Encord’s physical AI hub for practical advice and insights from industry leaders.

  • Access expert talks, private roundtables, and peer networking opportunities

  • See how ML Leads and Heads of AI are innovating faster across their data pipelines

  • Get insights into building large-scale 3D & LiDAR annotation and multimodal data pipelines

Join the hub now

partner with us

TOP_SIGNALS

Anthropic adds memory to Claude for teams, with project-specific context and incognito chats

1,947 Likes

Cursor improves Tab, its inline code suggestion system, with RL for 28% higher acceptance

1,749 Likes

Hugging Face integrates Copilot Chat provider to bring open LLMs directly into VS Code

1,358 Likes

OpenAI and Microsoft sign non-binding MoU, affirming nonprofit’s $100B+ equity stake in PBC

4,148 Likes

VS Code’s new update adds improved agent integration, context handling, and Auto mode for model selection

1,058 Likes

TOP_READS

Anthropic shows how to prototype and refine agent tools

2,846 Likes

Anthropic outlines best practices for designing AI agent tools. It explains clear naming, structured inputs/outputs, context returns, and token-efficient responses, with examples using Claude to prototype, evaluate, and refine tools.

Google DeepMind explains GPU memory, networking, and collectives

1,738 Likes

Google DeepMind details how GPU compute, memory, and interconnect bandwidth constrain model scaling. It explains collective operations inside nodes and across clusters.

Hugging Face brings GPT-OSS tricks to Transformers

1,947 Likes

Hugging Face details new Transformer features: faster model loading, reduced memory use, attention sinks, sliding window/hybrid attention, and quantization tools for large-scale models.

TOP_TUTORIAL

Google tutorial shows how to build MCP servers with Gemini CLI

2,583 Likes

alpha_signal_image_4

The tutorial explains Model Context Protocol (MCP) and demonstrates how to integrate it with Gemini CLI. MCP is a standard that lets large language models communicate with external programs, extending their abilities beyond text generation.


Key steps:

  • Understanding MCP: What it is and why it matters

  • Building a server: Create MCP servers in JavaScript or Python

  • Gemini integration: Configure settings.json to connect servers

  • Using OAuth: Authenticate services like Linear for secure task management

  • Veo 3 connection: Generate videos from text prompts via MCP

You will learn how to extend Gemini CLI with MCP to fetch data, update workflows, and trigger actions programmatically.

Watch 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 250,000+ AI developers?

Work with us