Skip to main content
Back

Qwen3-Coder-Flash is here (Open Source)

Gemini 2.5 Deep Think; Planetary Mapping; FLUX.1; Wide Research; Horizon Alpha; Higgs Audio V2; Kimi

August 2, 2025 3 min read

Hey,

Welcome to AlphaSignal, the most read source of news by AI engineers and researchers. Every day, we identity 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: 3 min 56 sec

Top News

▪ Alibaba Releases Open-source Qwen3 Coder Flash

Vellum

▪ One platform to build, test, and ship agents you’ll actually use

Trending Signals

▪ GoogleDeepMind releases Gemini 2.5 Deep Think

▪ BFL launches FLUX.1 Krea open-weights photorealistic image model

▪ ManusAI launches Wide Research for handling large documents

▪ OpenAI-linked stealth model sets new benchmark records on OpenRouter

▪ GoogleDeepMind launches a Unified AI Model for Planetary Mapping

Top Models

▪ higgs-audio-v2-generation-3B-base: Text-to-speech model, excels at expressive multilingual synthesis

▪ Kimi-K2-Instruct: Mixture-of-experts language model, strong autonomous code generation

▪ Qwen3-30B-A3B-Instruct-2507-GGUF: Language model, improved instruction following and 256K context

Top Tutorial

▪ Coding Qwen3 from scratch (notebook included)

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

It helps us keep this content free.

./TOP_NEWS

Alibaba Releases Open-source Qwen3 Coder Flash

3,181 Likes

Alibaba unveils Qwen3-Coder-Flash, a 30.5B code model that matches Claude Sonnet 4 on key coding tasks. The model runs on standard consumer hardware, setting a new bar for local deployment. 


Architecture and Technical Specifications
Qwen3-Coder-Flash features a Mixture-of-Experts architecture for efficiency and performance.

  • 128 experts, 8 activated per inference, with 3.3B parameters active at a time

  • 256K native context window, expandable to 1M tokens using YaRN

  • Full-precision inference requires 33GB of RAM; quantized versions run on 18GB

  • Model file size is about 60GB for download and storage

Benchmarks and Performance
Qwen3-Coder-Flash delivers strong results on several technical benchmarks for code and agent tasks.

  • Scores 51.6% on SWE-bench Verified for code reasoning and completion

  • Achieves 51.0% on Mind2Web for agentic browser tasks

  • Performs competitively on WebArena and BFCL for tool use and workflow integration

Licensing and Compatibility
The model is fully open source under the Apache 2.0 license and integrates with popular tooling.

  • Weights are available on Hugging Face and Modelscope for unrestricted use

  • Compatible with Hugging Face Transformers, LM Studio, Ollama, llama.cpp, MLX-LM, and KTransformers

  • Supports agentic coding and function calling with OpenAI-compatible APIs

  • Documentation and quickstart guides are available on Hugging Face and GitHub

Try now

One platform to build, test, and ship agents you’ll actually use.

Most agent frameworks bury you in complexity and edge cases you’ll never use.


Vellum cuts through the noise, it gives you direct control over what matters: context, memory, and integrations.

  • Decide exactly what your agent sees and uses

  • Load only the docs and tools needed for each task

  • Connect to 1,000+ platforms including Slack, Linear, HubSpot

  • Trace, evaluate, and debug context issues in real time

  • Deploy anywhere with a single click

Build agents with context you can see and trust.

Try vellum

partner with us

./TOP_SIGNALS

GoogleDeepMind releases Gemini 2.5 Deep Think for complex problem-solving

4,994 Likes

GoogleDeepMind launches a Unified AI Model for Planetary Mapping

3,403 Likes

BFL launches FLUX.1 Krea open-weights photorealistic image model

2,775 Likes

ManusAI launches Wide Research for handling large documents

2048 Likes

Horizon Alpha: OpenAI-linked stealth model sets new benchmark records on OpenRouter

1,796 Likes

./TOP_MODELS

higgs-audio-v2-generation-3B-base

157,021 Downloads

Higgs Audio V2 helps you generate expressive, multilingual speech with high-quality prosody and zero-shot TTS. It performs well on benchmarks like Seed-TTS Eval (WER 2.44, SIM 67.7%) and ESD (WER 1.78, SIM 86.1%). The model uses a 25 FPS audio tokenizer and a DualFFN architecture.

Kimi-K2-Instruct

352,110 Downloads

Kimi K2-Instruct is a 32B-activated parameter mixture-of-experts (MoE) model designed for advanced reasoning, coding, and tool use. It handles tasks like autonomous problem-solving and code generation, achieving 53.7% pass@1 on LiveCodeBench and 65.8% accuracy on agentic coding tasks. Trained on 15.5T tokens, it uses the Muon optimizer for stability.

Qwen3-30B-A3B-Instruct-2507-GGUF 

94,192 Downloads

Qwen3-30B-A3B-Instruct-2507 is an updated version of the Qwen3-30B-A3B non-thinking mode, with improvements in instruction following, reasoning, text comprehension, coding, and tool usage. It boosts long-tail knowledge across languages, aligns better with user preferences, and supports 256K long-context.

./TUTORIAL

Coding Qwen3 from scratch (notebook included)

Sebastian Raschka’s “Qwen3 From Scratch” tutorial provides a step-by-step guide to implementing the Qwen3 language model family using PyTorch.


The tutorial covers building the Qwen3 0.6B model, downloading pretrained weights, initializing the tokenizer, and running text generation.


Raschka compares Qwen3’s architecture to Llama 3, explaining trade-offs between depth and width, and details performance metrics for different hardware and model sizes.


The guide demonstrates use of the reasoning-focused model variant and offers tips for speeding up inference, including compilation, KV cache, and batched inference.


All code and instructions are available in the provided Jupyter notebook and repository.

Read 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 200,000+ AI developers? Let's work together.

Work with us