Alibaba Qwen3.5 Challenges OpenAI Dominance in AI Race
Open-weight AI with 17B active parameters delivers 60% cost savings and 8x throughput—redefining what free models can do
In-depth coverage, analysis, and updates on Open Source in AI and tech. 21 articles on AI Pulse.
Open-weight AI with 17B active parameters delivers 60% cost savings and 8x throughput—redefining what free models can do
The French AI startup is challenging OpenAI and Google with a radical strategy: give away your best work, then charge for expertise. It's either genius or financial suicide.
Meta's newest open-weight model sets new records across reasoning, coding, and multilingual tasks, intensifying the open vs. closed AI debate.
The latest open-weight model from Google challenges Meta's Llama and raises the bar for what's freely available to developers.
CropWise AI, trained on 15 years of African agricultural data, gives smallholder farmers personalized planting, irrigation, and pest management advice via basic SMS.
From autonomous hackers to on-device multimodal models, these repos are defining the cutting edge. Combined: 800K+ stars.
From automating 4,000 emails overnight to building custom trading bots, the open-source AI agent has 135K GitHub stars and a thriving ecosystem. We compiled the most useful repositories and creative use cases.
Meta's open-weights model outperforms OpenAI's flagship on HumanEval and MATH benchmarks. Anyone can run it locally.
A team of college students created an app that uses your phone camera to translate ASL to text instantly, and they open-sourced the whole thing.
Internal documents show Llama 4 outperforming all competitors. Meta plans to release it open-source in March.
Mistral Large 3 matches GPT-4o on most benchmarks and runs on a single GPU. The open-source AI movement just got a major boost.
China's MIT-licensed model with 685 billion parameters matches frontier performance. OpenAI responded by open-sourcing their own models.
Llama 4 Scout and Maverick are out now. Behemoth — with over a trillion parameters — is coming. Llama has hit 1 billion downloads.
Llama, DeepSeek, Mistral, and Qwen are catching up to closed models. The implications are enormous.
Moonshot AI's 1 trillion parameter model is open source, multimodal, and running 100 agents in parallel.
The French AI lab releases their most powerful model yet. Running locally is now viable for serious work.
Trained for $8M. Matches GPT-5 on reasoning. Open-weights. The cost efficiency gap with US labs is getting embarrassing.
405B parameters, state-of-the-art performance, and completely open weights. Zuckerberg's strategy is working.
China's open-source AI lab spent 1/15th what OpenAI did. The 'Sputnik moment' is real.
The open-source AI leader adds native multimodal understanding. Commercial use allowed.
The latest open-weight model matches proprietary alternatives on most benchmarks while remaining freely available