1. Qwen 3.8 27B ↗

A compact 27B native vision-language model for coding, professional work, research, and long-horizon agents, distributed as FP8 weights compatible with Transformers, vLLM, SGLang, and TokenSpeed. Its deployment-friendly open weights, near-original FP8 performance, multimodal support, and strong coding and computer-use benchmarks make the new model technically compelling to Hacker News readers. — 📊 Traction signals: Hugging Face shows 42 likes and 96.7k Qwen followers. Benchmarks include 61.7 on SWE-bench Pro, 79.0 on QwenSWEBench, 84.3 on OSWorld-Verified, and 90.3 on LiveCodeBench v6.

▲ 1182 · 💬 706 · ⚡ 74.5/hr

2. GLM-5.3 ↗

An AI model focused on complex coding, long-horizon agent tasks, and vulnerability discovery, improved over GLM-5.2 entirely through scaled post-training; open weights are promised two weeks after launch. A fresh model release combining frontier coding claims with unexpectedly strong cyber capabilities—and a quantified real-world vulnerability disclosure effort—creates both technical excitement and safety debate. — 📊 Traction signals: 50% improvement over GLM-5.2 on Z.ai Code Bench; Terminal Bench 3.0 rose from 4.6 to 28.3; CyberGym scored 84.5%; ExploitBench reached 54.4%; 2,436 vulnerabilities tracked across 269 projects, including 1,097 critical and high-severity findings.

▲ 1100 · 💬 543 · ⚡ 45.3/hr

3. RustDesk ↗

RustDesk’s preview build enables unattended Wayland remote access, including multi-monitor operation and login-screen connections after reboot, on x86_64 Debian/Ubuntu-based systems. It addresses a persistent Linux remote-desktop limitation that major competitors still handle poorly. The practical breakthrough drove 304 points, 129 comments, and 19 points per hour.

▲ 304 · 💬 129 · ⚡ 19/hr

4. AI by Hand ↗

Prof. Tom Yeh’s publication teaches AI mathematics, algorithms, and architectures through articles, interactive lessons, and hands-on workbooks covering topics such as activations, matrix multiplication, and self-attention. Its tactile, educational alternative to black-box AI content resonated broadly, earning 317 points at 16.7 points per hour; only 23 comments suggest appreciation outweighed controversy. — 📊 Traction signals: Over 73,000 subscribers; the completed activation chapter contains 12 interactive lessons.

▲ 317 · 💬 23 · ⚡ 16.7/hr

5. eigendrum ↗

An interactive simulation where users draw a drumhead shape, strike it at different points, and hear overtones computed from the shape’s permitted standing-wave modes. The unusual combination of mathematics, physics, audio, and direct browser interaction attracted 171 points, 45 comments, and 14.5 points per hour.

▲ 171 · 💬 45 · ⚡ 14.5/hr

6. Ploopy A+ Trackball ↗

An open-source, QMK-powered ambidextrous trackball kit with eight buttons, two programmable knobs, gestures, layers, on-device configuration, and a detachable wrist rest. Its repairable, hackable hardware design and substantial feature upgrade appealed to input-device enthusiasts, generating 109 points, 56 comments, and 12.8 points per hour. — 📊 Traction signals: Preorders open August 19, 2026 at 10am ET for $99 CAD; eight buttons, two knobs, eight-direction gestures, a Pixart PMW-3360 sensor, and a 1,000Hz polling rate.

▲ 109 · 💬 56 · ⚡ 12.8/hr