Open Source
Primary audience: Practitioners + vendors. Paired Signal dashboard: Hugging Face / OpenRouter Observatory.
Chinese Open Models Haven't Just Won the Leaderboard — They're Now Pricing the Hyperscalers
Three months after Chinese labs swept the top of OpenRouter's usage rankings, the lineup has barely changed. Moonshot is using that leverage to negotiate a cut of Microsoft, Amazon, and Google's hosting revenue — a reversal open-weight economics weren't supposed to produce.
Meta's Giant Muse Spark Finally Goes Open — With the Old Llama Strings Reattached
The 816-billion-parameter model Mark Zuckerberg promised to open-source in August landed on Hugging Face this week under a commercial license clause that Meta's smaller Muse Glimmer release had specifically dropped two months earlier.
More in Open Source
Nemotron 3.5 Bolt's native 2-bit checkpoint doesn't load in llama.cpp or Ollama, so a third-party GGUF repack is doing the work of actually getting it onto practitioners' machines.
pplx-decider, a 26B classifier fine-tuned from Qwen3.8-27B, showed up on Ollama this week with almost no fanfare. A third-party leaderboard already puts its newest version at the top of the growing field of “typed decision models.”
While Reflection AI and Mistral tease open-weight models still waiting on their public release, a Shanghai AI Lab-led team quietly shipped the whole stack for FlashVSR+ — paper, code, weights, and training data — under Apache 2.0.
A dataset behind an academic hallucination benchmark jumped 108% in daily downloads this week. The leaderboard it feeds hasn't been touched since 2024 and is currently throwing a server error — which makes the spike harder to explain, not easier.
StepFun's Step-5 Preview quadrupled its token share this week to crack OpenRouter's top 15, undercutting GPT-5.6 Sol on price while matching its benchmark score — but the open-weight release the company has promised is still five days away, and unconfirmed.
A week after launching d1 as a hosted API that reportedly undercuts GPT-6.1 Sol and Claude Opus 5.5 by up to 200x on narrow classification tasks, Liquid AI released smaller open-weight versions built to run directly on phones, cameras and robots.
The 740-million-parameter model maps text, code, images, video, and audio into a single vector space and runs on a Pixel phone in under 600 MB, extending the original EmbeddingGemma's 20-million-download run beyond text-only retrieval.
AutoTrust AI, a Singapore-based lab few outside the open-weights community had heard of two weeks ago, shipped three open 'decision model' releases in six days — and today's download and likes data show real, sustained pickup rather than a one-day launch spike.
PTB-XL, a cardiology dataset first published five years ago, saw four independent Hugging Face re-hosts jump in downloads on the same day — a pattern that says more about how open medical data circulates than about any single new paper.