Skip to main content
Back

Alibaba drops open Qwen3-Omni, first true end-to-end omni-modal

Sam Altman's blog, Figma’s MCP server, Open DeepSeek V3 variant, Google's Concept Design tool, OpenAI’s Stargate sites

September 24, 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 31 sec

Top News

▸ Alibaba rolls out Qwen3 updates, led by open-source end-to-end omni-modal AI

Voxel51

▸ Uncover hidden model failures using Voxel51’s scenario-driven analysis

Signals

▸ Sam Altman details his blog OpenAI plan to build gigawatt-scale AI capacity

▸ Figma releases remote MCP server, linking design data to agentic tools

▸ DeepSeek open-sources V3 variant, improves language and tool use

▸ Google Labs launches Mixboard, an AI board for visual concepts

▸ OpenAI, Oracle, SoftBank expand Stargate with five new data centers

Trending Repos

▸ every-programmer-should-know: Curates core CS topics you should revisit

▸ tldraw: Extendable framework for building collaborative visual interaction interfaces

▸ ROMA: Provides base architecture for building complex agent workflows

Top Read

▸ OpenAI on why it build Responses API for agents

Top News

Alibaba introduces open-source Qwen3-Omni, a unified model for text, images, audio, video

alpha_signal_image_2

Alibaba unveiled Qwen3-Omni, its first natively end-to-end omni-modal model that processes text, images, audio, and video in a single architecture.


Unlike earlier cross-modal systems, Qwen3-Omni trains jointly across modalities without degrading performance, and it supports real-time streaming outputs in both text and natural speech.


Key Features

  • Processes text, images, audio, and video in one model

  • State-of-the-art on 32 of 36 audio and audiovisual benchmarks

  • Multilingual support: 119 languages (text), 19 (speech recognition)

  • Real-time streaming with latency as low as 211ms (audio-only)

  • Handles up to 30 minutes of audio input

  • Customizable style and behavior via system prompts

  • Tool calling for external service integration

  • Includes an open-source universal audio captioner

How It Works

  • Thinker–Talker: Thinker generates text; Talker streams speech in real time.

  • AuT encoder: Trained on 20M+ hours for strong audio representation.

  • MoE: Sparse mixture boosts efficiency and concurrency.

  • Multi-Codebook: Step-by-step codec frames for low-latency speech.

  • Joint pretraining: No performance loss across modalities.

Performance

  • Matches same-sized Qwen single-modal models on text tasks

  • Outperforms closed-source Gemini-2.5 Pro and GPT-4o-Transcribe in speech recognition and transcription

  • Open-source SOTA on 32 benchmarks, overall SOTA on 22 audio and audio-visual tasks

Other Qwen3 Releases

Alongside Omni, Alibaba also announced:

  • Qwen3-VL: Flagship vision-language model (open-sourced) with GUI agents and visual coding.

  • Qwen3-Max: New Instruct/Thinking models, strong on coding and agentic performance.

  • Qwen3-LiveTranslate-Flash: Real-time interpreter with 3s latency and gesture/lip reading.

Try now
alpha_signal_image_3

Uncover hidden data issues with Voxel51’s scenario-driven AI model workflows

Tesla’s AI missing pedestrians is a clear reminder that high-performing models can hide critical data issues that silently degrade model reliability.


Voxel51’s Scenario Analysis, a customizable model evaluation workflow, helps ML teams go beyond surface-level checks to uncover hidden issues early. 


Deeply understand your multimodal data to discover:

  • Missing edge cases in training data that models fail to handle

  • Annotation mistakes that teach the model wrong associations

  • Data bias that skews model predictions

  • Poor quality samples that cause overfitting

Read the blog

partner with us

Signals

Sam Altman writes in blog about OpenAI's new financing strategies for upcoming large-scale AI infrastructure projects

4,847 Likes

Figma adds remote MCP server, letting Claude Code and agentic tools directly access live design context

9,736 Likes

DeepSeek launches open-source V3.1-Terminus, improving language, coding, and search performance significantly

4,582 Likes

Google debuts Mixboard, an AI tool that combines text prompts and image editing to support concept design

3,922 Likes

OpenAI, Oracle, and SoftBank add five Stargate sites, advancing toward $500B 10GW infrastructure target

4,119 Likes

Repos

every-programmer-should-know

91,291 Stars

It collects foundational CS topics, “falsehoods,” architecture principles, and more. You can use it as a reference to fill gaps and structure your learning in engineering.

tldraw

42,880 Stars

This repo offers an infinite-canvas whiteboard SDK in React. It includes context helpers (CONTEXT.md) to help LLM agents navigate and understand code structure. Useful as a visual tool in AI systems.

ROMA

3,411 Stars

You get a meta-agent platform where you design agents that reason across levels: plan, execute, and aggregate. ROMA handles recursive control flow, transparency, and tool wiring. Good for long-horizon tasks.

Top Read

OpenAI on why it build Responses API for agents

1,229 Likes

alpha_signal_image_4

OpenAI built the Responses API as the next step beyond Chat Completions, designed for reasoning-heavy, tool-driven, multimodal workflows. 


You’ll see:

  • How /v1/responses differs from Chat Completions

  • How reasoning state persists across calls (+5% on TAUBench)

  • How structured outputs reveal tool calls, receipts, and intermediate steps

  • How hosted tools like code interpreter, web search, image gen plug in directly

  • How it improves cache efficiency, latency, and cost

  • How it handles reasoning safely without exposing raw chain-of-thought

If you care about building agents that don’t drop context and can run safely at scale, this blog is worth reading

Read 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