I reverse engineered Qualcomm's NPU compiler to find undocumented behaviour that affects every edge AI deployment.
Things nobody knew:
1. The compiler silently downgrades the precision of your model weights without telling you
2. Memory placement uses HiGHS which is an LP solver (not heuristics)
3. The same model on two different chips with identical reported VTCM can have 33x difference in DDR traffic
4. There's an undocumented internal simulator called Hextimate pricing ops without the hardware
Every NPU vendor be it Qualcomm, MediaTek or Apple NEVER tells you how to make the most use of their hardware.
I was very close to rage quitting before I finally lost all hope and thought of reverse engineering to understand how NPUs are handled.
Read the full write-up below:
We built the fastest PDF -> markdown parser in the world 🚀⚡️
AND it’s more accurate than any other open-source, model-free parser (pymupdf4llm, opendataloader, pdf-inspector, markitdown) on 3 standardized benchmarks: olmOCR0-bench, opendataloader-bench, ParseBench
Introducing LiteParse v2.1. The v2 base version was already the fastest document->text parser on the planet, and with this new release we’ve introduced markdown.
It is fully open-source (Apache 2.0) and free, is usable from CLI/Rust/Node/Python/WASM, and is also installable as a one-click agent skill.
Check it out: https://t.co/7oFImAZeb2
Come check out LiteParse: https://t.co/JNER0mVcB8
This panda is just hanging out at National Zoo & Aquarium Canberra 🐾
🌳 As an arboreal species, red pandas spend most of their lives in the trees—making healthy forests essential to their survival.
⚠️ Unfortunately, red panda habitat has been fragmented by agricultural expansion and human settlement, leaving their remaining home divided into hundreds of isolated forest patches. This fragmentation threatens the long-term survival of red pandas and other wildlife.
That’s why we’re working alongside local communities 🤝 to restore and reconnect forests through our #PlantARedPandaHome campaign—creating a community-protected biological corridor where red pandas and other threatened species can thrive.
🙏 We’re grateful to partners like National Zoo & Aquarium Canberra for supporting this critical work through their annual donations.
Learn more about habitat restoration and how you can help: https://t.co/VUABJE9cTV
#RedPanda #HabitatRestoration #WildlifeConservation #CommunityConservation #Reforestation #ProtectWildlife #SaveTheRedPanda #NationalZoo #Nepal #ThankfulThursday
It is the first open-weight model to break 30.0% on ProofBench, scoring 11.0% higher than the second-ranked model. It also outperforms Gemini 3.5 Flash on ProofBench and trails Opus 4.5 by only 1.0%
Today I saw @Esychology et al's demo of high-res neural cellular automata and it was cool - ever since seeing NCA on distill in 2020, always wanted to implement one
Here is their browser demo ported to jax-js - no custom shaders, just jit(), ~100 FPS :D
Same code runs on both WebGPU + Wasm
https://t.co/gT9G62jhMH