GLM 5.3 just released which is supposedly good at security work and doesn’t simply throw up its hands and say “nope” like Anthropic. I’m running an audit on gRPCServer.jl as we are getting it ready for release, will have to see what it comes up with. If it pans out I’m thankful to finally have a model for this as an open source maintainer.
GLM 5.3 just released which is supposedly good
GLM 5.3 is a really strong open weight model, costs around half of sonnet if not a third and solves many things as well as if not better. I use it a lot with my startup getslaide.com - for the hosted version, simply because of pricing etc. (but it also proved to be better and more consistent in local testing)
Let us know the result! 
It found a DoS attack in gRPCServer.jl that could bring the server down in a single message, a correctness bug that could result in a hang/deadlock for response streaming in gRPCClient.jl, and found several bypasses allowing for arbitrary code execution in a plan mode plugin for Pi that I had been developing. Keep in mind I had already done many rounds of hardening in gRPCClient.jl so all low hanging fruit had been eliminated. gRPCServer.jl is much newer, so it hasn’t had nearly as much attention.
It was interesting seeing GLM 5.3 do its adversarial testing, coming up with all kinds of inputs to poke at different parts of the system. It did this in a way I havn’t seen Anthropic models be willing to do; they seem to just more or less only focus on the code itself and avoid anything that could lead to a reproducible exploit.
One more update, I’m running Qwen 3.8 27b 8bit locally on vLLM. This model is seriously impressive considering its size, it seems to actually be better than Deepseek V4 Flash in my experience. I had it build an a modern X11 terminal emulator focused on fast rendering without GPU acceleration, and it built it overnight from like 8 different design requirements. A few rounds later I have full TUI applications running inside it (herdr + pi), resize working properly, etc. I’ve also been using it to make improvements for DocumenterLandingPage.jl and it’s been doing a great job there as well.
The NVIDIA research project Agentic Variation Operators (AVO) addresses this challenge of building a general-purpose agent architecture. AVO was first demonstrated on difficult software engineering and GPU-kernel optimization tasks, where success requires far more than generating code in a single response.
..
GPU-kernel optimization and the ARC-AGI-3 benchmark appear very different on the surface.One involves source code, compilers, profilers, and throughput. The other involves unfamiliar interactive environments in which an agent must infer the effects of available actions, discover objectives, and act efficiently enough to make progress.
But the underlying computational pattern is similar.
..
Using Claude Opus 5, AVO completed the full 25-environment public set with a 100.00 RHAE score, solving all 183 levels in 6,624 environment actions.
..
The most important result was not simply the 100.00 score, but that the same agent architecture transferred from highly specialized GPU-kernel optimization to a very different interactive reasoning task.
Ox Alpha model is claimed frontier, and (still) free to use (and most used with DeepSeek harness at OpenRouter):
It’s most used on all the harnesses right now or near the top, with e.g. Claude Code (most used with it last few days): DeepSeek Harness | OpenRouter and App & Agent Rankings | OpenRouter
EDIT: I managed to use this model a bit for free, coding Julia, with Deepseek harness through openrouter, until I got: This turn failed 404: {"message":"Thank you for participating in the Stealth Ox Alpha testing period. This model was ZAI's GLM-5.3 Flash. Use it now: https://openrouter.ai/z-ai/glm-5.3-flash","code":404}