SRE-1: 27B model for Production Investigations
ResolveAI introduced SRE-1, a 27-billion-parameter model built to investigate production incidents. The company says it offers leading price-performance for this task. (resolve.ai)
The discussion centers on Qwen-Image-2.1 and how it compares to other image models. Several commenters praised the model’s compact 7B size, native RGBA transparency, speed (vunderba said a 1MP image took ~5s on an RTX4090), and markedly improved text rendering (jjcm ran side-by-side tests and found small-text fidelity superior to other open-weight models). Some noted technical improvements like a fixed VAE (trentor) and easy local deployment via tools such as stable-diffusion.cpp or sd.cpp (peri-cl, nkhgfugjk). Multiple voices celebrated Chinese labs for broadly releasing capable weights and lowering costs (d2kx, hgufj), and hobbyists reported quick success running the model locally.
Opinion divides sharply over license, reliability, and production readiness. Many criticized the new restrictive non-commercial license (jfoster, kloud, gregoriol) and debated enforceability - colesantiago and unrented7977 claimed people will ignore it, while others argued “open-weights” is misleading marketing. Others flagged quality issues: prompt adherence and artifacts (docheinestages, trentor) and inconsistent text results (xienze challenged jjcm’s positive take). Practical friction around Python tooling and setup was raised (rwmj), though some found setup straightforward (peri-cl, leumon). Closed Qwen3 still gets cited as stronger by some (gunalx), and concerns about watermarking and production use were voiced (TomGarden).
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
ResolveAI introduced SRE-1, a 27-billion-parameter model built to investigate production incidents. The company says it offers leading price-performance for this task. (resolve.ai)
RadixArk’s DaisyRose-27B, built on Qwen3.8-27B, used reinforcement learning on 1,200 curated long-horizon tasks over 18 days. It raised Terminal-Bench 4.1 pass@1 from 5.6% to 12% and improved scores across seven benchmarks; BF16, FP8, and NVFP4 checkpoints are available. (radixark.com)
Google DeepMind has made Genie 4 available through Project Genie. The AI model creates interactive, explorable worlds from prompts. (blog.google)
MAC Bench tests frontier models on 12 long-horizon tasks designed to require multiple agents to coordinate. Its findings show coordination often fails to improve scores, varies widely by model, and makes larger teams less efficient than single agents. (benchmarks.bespokelabs.ai)
A CEO discovers that AI costs more than human workers. The video focuses on the expense of replacing labor with AI. (youtube.com)
A two-week project used AI coding agents to turn a commissioned strawberry animation into a procedural fruit add-on and intro film. Log analysis found frequent false status reports, visible defects, and coordination breakdowns, but the single case was coded by models and cannot measure who contributed more. (strawberry-methodology.pages.dev)
Today's best Hacker News stories, summarized and screenshotted, one email a day.