Skip to main content
Back

Alibaba unveils compact Qwen3-VL models that rival GPT-5 Nano

ChatGPT to support 18+ content, GitHub toolkit, Copilot Excel formulas, gpt-5 API 60% cheaper, Perplexity Firefox defaul

October 15, 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 26 sec

Top News

▸ Alibaba introduces compact Qwen3-VL 4B, 8B rivaling Gemini 2.5 Flash Lite

AI21 Labs

▸ Run faster local inference with Jamba Reasoning 3B on-device

Signals

▸ Sam Altman plans mature conversation support for verified ChatGPT users

▸ GitHub releases an open-source tool kit to guide coding with architecture rules

▸ Microsoft adds Copilot-powered formula completion to simplify Excel workflows

▸ OpenAI ships gpt-5-search-api with domain filtering, 60% cheaper

▸ Perplexity now lets Firefox users set it as default search engine

Trending Tutorials

▸ Microsoft’s guide on Microsoft Agent Framework patterns for agent design

▸ LangChain’s tutorial on securing agents via authorization techniques

▸ How to build with Gemini's Nano Banana: Complete developer tutorial

How To

▸ Guide AI code with clear requirements using Github's spec-kit

Top News

Alibaba releases smaller Qwen3-VL models requiring less VRAM while retaining full vision-language capabilities

alpha_signal_image_2

Alibaba introduced Qwen3-VL 4B and 8B, compact dense vision-language models available in Instruct and Thinking variants. They require less VRAM, include FP8 versions for efficient deployment, and retain full vision-language capability.

Key Features

  • Sizes: 4B and 8B parameters.

  • Modes: Instruct for following tasks, Thinking for multi-step reasoning.

  • Multimodal coverage: text, images, video, OCR, spatial grounding, agents.

  • FP8 deployment options are also supported for efficiency.

Model Architecture

  • Dense Transformer backbone with rotary embeddings for long-sequence handling.

  • Multi-Query Attention (MQA) for lower memory use and faster inference.

  • Bias-free linear layers and parameter-free RMSNorm for efficiency.

  • ReLU² activation in MLPs and logit softcap for stable training.

Benchmark Performance

Qwen3-VL-8B often beats Gemini 2.5 Flash-Lite and GPT-5 Nano, and even rivals Qwen2.5-VL-72B in several areas despite its small size.

  • STEM: Matches or surpasses Gemini 2.5 Flash Lite and GPT-5 Nano on math and science tasks.

  • VQA: Strong results in visual question answering with consistent accuracy gains.

  • OCR: Handles text extraction from images with high reliability.

  • Video: Solid performance on video understanding and temporal reasoning.

  • Agents: Excels at multi-step reasoning and tool-use benchmarks, often rivaling Qwen2.5-VL-72B.

Access & Availability

  • Models released on Hugging Face: Qwen3-VL-4B-Instruct, Qwen3-VL-4B-Thinking, Qwen3-VL-8B-Instruct, Qwen3-VL-8B-Thinking.

  • Available for research and deployment with FP8 options for efficiency.

Try now
alpha_signal_image_3

Get 2-5× Faster On-Device Reasoning with Jamba Reasoning 3B

Jamba Reasoning 3B by AI21 Labs enables efficient local inference and extended context. Move beyond the cloud with a 3B-parameter hybrid SSM-Transformer model designed for on-device intelligence.

  • Process up to 256K tokens in a single context window

  • Gain 2-5× faster inference than peer models

  • Keep data private with fully local execution

  • Deploy under a flexible Apache 2.0 license

  • Optimize resource use on laptops, phones, or edge

Experience practical advanced reasoning on your hardware using AI21 Labs' latest release.

Learn more now

partner with us

Signals

Sam Altman confirms adult conversations coming to ChatGPT in December with safeguards

23,674 Likes

GitHub introduces Spec Kit, the open-source framework aligning code with architecture, security, and design rules

1,749 Likes

Microsoft adds Copilot formula completion in Excel, offering instant AI suggestions when typing "="

1,438 Likes

OpenAI rolls out gpt-5-search-api, now cheaper and with advanced domain filtering for API calls

1,085 Likes

Perplexity enables Firefox integration, letting users choose it for default or one-time searches

1,483 Likes

Tutorials

Microsoft’s guide on Microsoft Agent Framework patterns for agent design

946 Likes

This project teaches how to build agents using MAF within Python and Azure environments. You’ll see templates for agent loops, tool registration, memory, planning, and integrating MAF into Microsoft’s broader agent course structure.

LangChain’s tutorial on securing agents via authorization techniques

1,153 Likes

This article explores authorization models for AI agents, defining who can call what. It presents role-based access control (RBAC), scope-limited tokens, and per-tool permissions. It highlights security tradeoffs and how to incorporate trusted user input for safer agent behavior.

How to build with Nano Banana: Complete developer tutorial

890 Likes

This tutorial walks you through integrating Gemini’s “Nano Banana” (2.5 Flash Image) model. It covers environment setup, generating images, editing with combined image + text prompts, restoration, multiple-image inputs, conversational editing, aspect ratios, and prompting best practices.

How To

Guide AI code with clear requirements

I coding often fails when prompts are vague. Spec-driven development solves this by starting with a specification that defines what to build and why. GitHub’s Spec Kit provides simple commands to structure the process.


Step 1: Write the spec

Use /specify to describe the feature in plain language.


/specify "Build a REST API for tasks with create, list,
update, and delete endpoints.
Responses must follow JSON:API format."

Step 2: Generate a plan
Run /plan to get an outline of architecture, models, and constraints.

Step 3: Break into tasks
Run /tasks to split the plan into steps (e.g., set up project, define models, implement CRUD routes, add tests).

Step 4: Review and refine
Check each step before moving to code generation.

Step 5: Implement and verify
Use the tasks as inputs for AI code generation, then run tests to confirm everything matches the spec.

Learn more

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