Hermes Vs OpenClaw
Posted by Birdinhandandbush@reddit | LocalLLaMA | View on Reddit | 38 comments
I always prefer to try local first because I don't have the money to waste, so I started running OpenClaw and Qwen3.5:b for a while but I found the limits of the model being like longer complex tasks with multiple steps and I was getting frustrated, user error I know.
So I saw posts on Hermes and decided to go test it as well. Same Qwen model, and it's been purring like a kitten and getting things done, my faith in local renewed.
But I didn't want to give up on the claw and one subscription ended freeing up some cash so I went cloud Ollama and GLM 5, and suddenly OpenClaw found the engine it needed.
It pretty much one shotted two separate issues without a break.
So I have an excellent experience now with my Hermes agent running some jobs, absolutely free and just getting better with Qwen locally,
and at the same time I have the experience I was hoping for a few weeks ago with OpenClaw and the more powerful GLM 5 model.
Maybe it's the frameworks, the way they work.
Maybe it's just a learning curve and I got better with them.
I'd almost lost faith in the Qwen model until I plugged it into Hermes, but I almost lost faith in OpenClaw before I plugged in GLM 5.
SnooPaintings8639@reddit
I am forcing myself to use Hermes, due to all the positive comments I read, but damn, I feel much better working with my OpenClaw agent.
wouldacouldashoulda@reddit
For what it’s worth I suspect a ton of astroturfing on Hermes. It’s ubiquitous and generic, stuff like “hermes agent has the best harness” and “Hermes is really good”, while we’re at the stage that we really should know a little more.
LolliaSabina@reddit
I am 100% a real person and I've tried it and I really like it. Just feels less buggy and like it's easier to find and do the things I want to find or do.
My only frustration seems to be that I do have some issues with sessions timing out or dying. Nobody else seems to be complaining about this though, so maybe it's an issue with my set up
Whatforit1@reddit
I had the same issue, the docs mention setting streaming timeouts in env bars. It says that it auto increases timeouts if it's a local model, but it didn't work for me, probably because I have mine on a public DNS.
deejeycris@reddit
it's definitely overhyped by the AI influencer brigade, it's still half-baked. I tried it extensively then decided to switch to pi agent. I'd rather get something not baked at all but with a good core and than I'll just slop out my own custom agent system and iterate, than working on somebody's else's slop.
Birdinhandandbush@reddit (OP)
Yeah I wouldn't have heard of it or tried it except for reading the posts here
Birdinhandandbush@reddit (OP)
Claw has more skills, a bigger ecosystem and more users, I definitely think it's more likely to be the biggest player in the long term, but Hermes seems to have a tighter structure on getting tasks completed with less powerful models, it's definitely been a learning experience for me
Long_Live_Bevo_XV@reddit
I'd love to hear feedback from anyone who's used both OpenClaw and Hermes for more than a week each. I see lots of positive comments about Hermes's memory system, but not much about how well it holds up over time as it accumulates cruft. Understandable since Hermes is much newer.
CommonPurpose1969@reddit
Hermes and Qwen3.5 4B is a beast. OpenClaw doesn't work because it was designed for LLMs first, and that strategy fails with SLMs. Always.
Birdinhandandbush@reddit (OP)
4b? I was thinking you'd need bigger. It's running well with my 9b model. If I dropped it to the 4b it'd be crazy fast, I just didn't think it would work. Tell me more
CommonPurpose1969@reddit
It surprised me, too, that it did work. And how well it worked. It basically wrote and executed Python code that it picked up from skills shipped with Hermes!
Birdinhandandbush@reddit (OP)
I might test this, not the code part, but a 4b model on my 16gb card would mean I could do image generation in ComfyUI at the same time with another small image model, or play some games in theory
pop0ng@reddit
Report back if successful please
Birdinhandandbush@reddit (OP)
Success. It's faaaast. Like I'm not using it for coding, more research and creative projects and I am impressed. The speed, it's fantastic. I'm not relying on it's intelligence, usually I'll pull in context or provide documents, so dropping to 4b was less of a shock, but I tested it for the afternoon and I think it's excellent so far. I could run this on an older laptop, seriously 😳
pop0ng@reddit
Nice! Im going to try it in WSL
chaosraser@reddit
What are you doing with hermes Agent? I installed it with ollama and gemma4, but die simples Questions are not coming back. Or i have a wrong idea how to use a local hosted Agent.
celsowm@reddit
At least openclaw runs natively on windows
Birdinhandandbush@reddit (OP)
Natively? It can I guess, but it's best run inside the Linux wsl
celsowm@reddit
I hate wsl
Birdinhandandbush@reddit (OP)
I hear you, but it's a safer sandbox than running an agent directly on your open system
GrungeWerX@reddit
Did you try running open claw with Qwen 27b?
Birdinhandandbush@reddit (OP)
I have 16gb vram so it's a squeeze and I didn't want to go to q3 quants , maybe that's a me problem
sirnixalot94@reddit
The frameworks have a lot to do with it. I used OpenClaw for a few weeks with Qwen3.5:9b on an RTX 4080 and it was doing well until all of a sudden one morning it just decided I was supposed to be creating summaries of the news articles it was set up to do instead of doing it itself. It hummed along fine for about three weeks doing exactly what it was supposed to do up until that point. After that I had to intervene every morning to get the jobs to complete.
I switched gears and converted most all of my jobs to cron triggered Python scripts only calling the LLM through Ollama when needed. From then on everything has been great as far as my daily jobs go. I then used Claude to help me build my own agent for research (market analysis, AI related product launches, business potentials, etc.) and that has been working pretty well now for a couple weeks.
Ultimately I want to be able to do web application/site development on strictly local models. None of that seems to be possible with a small model on OpenClaw, Agent Zero, or Hermes that I found. They lose focus too easily. The farthest I got on that front was using Zed with the aforementioned qwen model. I got excited when it iterated through the creation of every single file for a site I asked it to build… and then crapped the bed at the last minute trying to fix a syntax error. I tried Genma4:e4b before that and it couldn’t figure out how to make a directory or write files. I have much higher hopes for this endeavor in the coming months once I buy a (possibly a couple) DGX Spark variants.
GrungeWerX@reddit
Maybe it’s Maybelline.
jacek2023@reddit
so the conclusion is you need to pay to use OpenClaw and you can use Hermes Agent locally?
Birdinhandandbush@reddit (OP)
For me with my 16gb 5060ti OpenClaw can be a chore, it just kept stopping or hitting a wall, again could have been user error. But Qwen absolutely nailed it in Hermes, maybe they're just keeping a tighter structure on the way the model works, but out of the box it's been great. But the larger GLM model, I just don't have the hardware to run it locally, but it's working brilliantly for me under OpenClaw. I think it might be a few more updates before claw is as tight as Hermes. But I also think claw has more skills, more potential, so ultimately will be the winner.
TheThoccnessMonster@reddit
I’m anecdotally seeing the exact same thing with qwen in Claw.
Birdinhandandbush@reddit (OP)
I'm definitely not the only one. First few days I was just amazed with it in OpenClaw, but the more complex the tasks, the more difficulty I had, and it just wasn't doing what I told it. One of the key problems was delegation, I'd specifically ask for sub agents to run tasks and come back to find Main had tried doing everything, again I know, probably my fault too, but GLM hasn't had that problem and spits out sub agents at ease, does everything really well, like I'm back to loving Claw. But the shock was watching Hermes chug along with the same Qwen model and not once have I had to fight with it , oh wait, I just remembered, it did completely wipe a config file which was funny, I had to go to Github and get a copy,. but hey nobodys perfect.
Anyway, we're all learning, some of us more than others.
TheThoccnessMonster@reddit
I think that OpenClaw’s script timeouts is something that has to be basically in its context on every prompt as it results in it making assumptions that it will get a call back to respond and never does. I also noted it doesn’t doublecheck for failures to things like editing cron via open gateway to correspond with things you put in heartbeat etc.
That said as other has noted, $20 goes a long way in Ollama for the moment. Faced with $200 for Claude max versus $200 per year for basically similar use of GLM 5.1 is a no brainer.
InstaMatic80@reddit
Using ollama cloud and telling you prefer local first is a nonsense imho
Birdinhandandbush@reddit (OP)
Qwen3.5:9b is local, running on my GPU. It's still driving my Hermes agent. It's performing great. For a lot of people testing the water I can definitely recommend it. I'm only using the ollama cloud model for OpenClaw because I can't run a 500gb model on consumer hardware.
SelectionCalm70@reddit
hermes agent has the best harness for open source models . worth trying
Birdinhandandbush@reddit (OP)
I've had a few teething problems but once you get your head around it I think Hermes could be a much easier regular driver for most people.
Mkengine@reddit
I haven't used either of them so far, what kind of problems do OpenClaw or Hermes agent actually solve? who is the target audience? do I benefit from using it as a Software Engineer or is it for people who can't code?
Birdinhandandbush@reddit (OP)
I'm doing absolute minimal coding, that's not what I'm using it for. I'm a technical trainer and content creator, it's gathering information for me, meticulously cataloguing sources, that's vital for me, producing first drafts after that. It's a great record keeper, I can brainstorm with it and it can go off researching and drafting and let me know when to check output. It does a daily briefing for me. It's like having a junior employee. I use it for company finance keeping too. I photograph a receipt and it knows to read the numbers and where to store them. If I need a service I can either look at skills or ask it to build around what I need. It really depends on you
rosaccord@reddit
Absolutely agree.
I have just started running hermes with qwen 3.5 27b
really like it
Birdinhandandbush@reddit (OP)
I could try a q4 quant but it doesn't leave much room for context on my 5060ti. Maybe if we get a turboquant version
No_Conversation9561@reddit
Hermes is really good. I use it with Qwen3.5 27B.