5 stories that day · 1 model · 2 open source · 1 research · 1 trending
Published daily at 7 PM ET · Updated · RSS
ModelTop story
Only 17 billion parameters work on each step. It is on the lab's Tinker platform now, and the lab says it plans to share the full weights next week, once outside safety testing is done.
Thinking Machines Lab · Original source (opens in a new tab)
Open source
Tencent releases DSpine, a small draft model that helps Qwen3.8-27B write its answers faster (opens the original source in a new tab)DSpine drafts 16 tokens (word pieces) at a time and the main model checks them. In Tencent's agent tests, output came 2.71 to 3.84 times faster than normal decoding. It is under Apache 2.0.
Tencent · Original source (opens in a new tab)
Open source
Underdog releases Woof, a 40.6 MB text-to-speech model that runs on a laptop with no GPU or server (opens the original source in a new tab)It has two English voices and reads phone numbers, dates and prices as a person would. Underdog says it runs 13.8 times faster than real time on one CPU thread. Use is non-commercial only.
Underdog · Original source (opens in a new tab)
Trending
llama-modchip runs a 284-billion-parameter DeepSeek model on a computer with 64 GB of memory and no graphics card (opens the original source in a new tab)These open-source llama.cpp patches shrink the model to 56.56 GiB so it fits in memory, and only 13.17 billion parameters work on each step. The author reports 9.1 to 14.7 tokens (word pieces) a second.
343 points on Hacker News
synchopate · Original source (opens in a new tab)
Research
Each new chunk of video can draw on earlier chunks before they are finished, so the GPUs work side by side. The authors say quality stays on par with an earlier method, Rolling Forcing.
125 upvotes
Notable stories from the past few days. Each shows its own date.
