Hey everyone,
Been running Hermes on a cheap VPS for a couple of months now, mostly just firing prompts at it through Telegram when I’m away from my desk.
I was on DeepSeek for a solid run because the raw token speed is crazy good, but man, the tool calling was driving me nuts. The second I tried chaining browser tasks with vision models, things would just fall apart or hang indefinitely. You get this blazing fast response, but it just completely ignores the JSON schema or hallucinates the tool output halfway through. Super frustrating when you just want a simple scraper or screenshot check to work.
Swapped my backend over to GLM-5.3 Flash earlier today to see if it could handle function calls any better. Honestly, first impressions are way better than I expected.
It feels genuinely fast on Telegram, almost zero delay on the first chunk, and the tool calling actually holds up for short routines. I threw some basic file operations, API triggers, and light Python script generation at it, and it hasn’t choked yet. Usually these “Flash” optimized models lobotomize the reasoning just to hit high tokens-per-second benchmarks, but this one doesn’t feel like a stripped-down toy so far.
Obviously it’s only been a few hours, so I’m not declaring it a silver bullet yet. Over the weekend I’m planning to push it through some nasty edge cases: letting it run longer autonomous loops without supervision, throwing 5+ chained browser/vision tools at it in a single run, and intentionally feeding it broken inputs to see if it can recover gracefully or if it enters a death spiral of failed retries.
Anyone else running GLM-5.3 Flash as their daily driver backend for agents? How does it hold up once context gets messy, and are you guys tweaking the system prompts or just running default parameters?