Softtechhub

Drex Max v1 is Nace AI’s open-weight 27B decision model: typed questions in, probabilities out, one forward pass, on a 24 GB Mac. What the data really shows.

qwenGPU compiles a 4-lane SIMT GPU into Qwen3.5-2B weights, running real CUDA in forward passes. How it works, what it proves, and why it stays a research demo.

Auto Ontology is NVIDIA’s open semantic layer that grounds agent SQL in certified business definitions. How it works, and what its missing benchmarks mean.

Zyphra Hierarchical Attention (ZHA) searches a tree of compressed keys instead of the whole KV cache, retrofitted onto a frozen 35B model in a short distillation. What the data really shows

Pando, OpenAI’s new Git backend, cut median push latency by nearly 80 percent and serves about a petabyte of Git data a day. How it works and what comes next.

Soniox STT v6 streams the agent’s words back to the recognizer and filters its playback out of transcripts. How it works and what it changes for voice agents.

FlashProve is Anthropic’s open zkSNARK prototype for provable inference: proof that a committed model answered, at 110 tokens per second on one H100.

WeVisOne-VL is Tencent’s open 2B and 4B vision language model covering detection, OCR, counting, segmentation and matting. What the numbers really show.

A cloud computer for AI agents shares your Chrome logins, shows every action live and keeps sessions in your own drive. Where it helps, where it bites.