Research library
Software engineering
Practical architecture and development decisions for software that people can use, maintain, and adapt.
89 articles · Page 4 of 8
Tencent's Hy4 Preview Is a 770B Apache Model With an API-Sized Deployment Problem
Tencent Hy4 opens 770B-parameter weights under Apache 2.0 and claims strong agent performance, but deployment remains data-center scale.
GLM 5.3 Is a Two-Model Family With a Split License Strategy
GLM 5.3 pairs a 753B coding model with an MIT-licensed multimodal Flash model. The name is shared; the architecture and commercial terms are not.
Qwen3.8-Flash-Next Is a Qwen4 Architecture Preview. Its License Is the Catch.
Qwen3.8-Flash-Next activates 6B of 125B parameters and previews Qwen4. Its efficiency is notable; its commercial assistant license is restrictive.
How AI Watermarking Works: Text, Images, Audio, and What the Mark Actually Proves
How AI watermarks work in text, images, audio and video, what C2PA adds, what Article 50 and California's SB 942 require, and what a detected mark proves.
Cursor Launched a GitHub Rival on the Day GitHub Fell Over
Cursor's Origin puts repos, PRs and CI inside the editor, three days after SpaceX closed its $60B deal and on the day GitHub degraded for 7.5 hours.
A Month In, LM Studio Bionic's Cloud Charges What the Labs Charge. Its Subscription Is Still 'Coming Soon.'
LM Studio Bionic launched July 16 as an agent for open models. A month and eight releases in: what runs free, what the cloud costs, and what is unresolved.
Anthropic Flipped Claude Code to Auto Mode Because You Were Approving 97% of Prompts Anyway
Claude Code auto mode is now the default on Pro, Max and Team. Its study says humans caught 13.6% of bad commands. What to audit before trusting it.
DeepSeek Open-Sourced the Harness That Graded Its Own Benchmarks
DeepSeek Harness (dsh) is MIT-licensed and model-agnostic. It is also the software that produced V4-Pro's Terminal-Bench and DeepSWE scores.
OpenAI's Daybreak Red Ships a Model That Answers 95% of Exploit Requests. Blue Gets 2%.
OpenAI split Daybreak into Blue and Red tiers and gated GPT-5.6-Cyber behind Red. The gap between the tiers says where model refusals actually live.
Meta's Muse Code Cuts Your Token Bill 90%. The Currency Is Your Repository.
Muse Code ships with a contributor tier that cuts token prices over 90% if Meta can train on your code. What that trade is actually worth per developer.
DeepSeek Retrained a 284B Model Until It Beat Its Own 1.6-Trillion Flagship
DeepSeek's official V4-Flash keeps its April architecture and now beats V4-Pro-Preview on all nine agent benchmarks. Post-training did all of it.
Claude Opus 5: Anthropic Undercuts Its Own Flagship
Claude Opus 5 ships at $5/$25, half of Fable 5, with near-Fable scores, looser cyber classifiers and Anthropic's lowest misalignment audit score yet.