Same servant, different masters — er, runtimes
The Experiment
Vivek ran the exact same 4B parameter model (Qwen/Qwen3.5-4B) via LM Studio on his AMD Ryzen 5 5500 + RTX 3050 (6GB) home PC in Bengaluru. The only variable? The agent framework: OpenClaw for the first two days, then Hermes for the next five. The AI persona was "Suppandi" — a witty, sharp servant with a 🕺 emoji.
87
total messages
22 from Vivek · 65 from Suppandi
2 days (Mar 29–30)
VS
220
total messages
60 from Vivek · 160 from Suppandi
5 days (Apr 1–5)
Time from "user hits enter" to "first meaningful bot response"
Overall Response Latency
| Metric | OpenClaw | Hermes | Winner |
|---|
| Mean latency (raw) | 52.2s | 25.2s | Hermes |
| Mean latency (cleaned*) | 52.2s | 18.8s | Hermes |
| Median latency (cleaned*) | 29.5s | 15.0s | Hermes |
| Fastest response | 1.8s | 1.9s | Tie 🤝 |
*Cleaned = excluding model-loading failures (LM Studio model wasn't preloaded) and runaway context crashes, which are infrastructure issues, not framework speed.
Hermes is nearly 3x faster in practice. Once you strip out the outliers caused by LM Studio's model not being preloaded (which caused 218s and 303s spikes on Apr 4), Hermes clocks a clean 18.8s mean / 15.0s median. OpenClaw's 52.2s mean is inflated by constant context compaction cycles that add 10–91 seconds of overhead per compaction — and it compacted 6 times in 2 days.
Latency by Query Complexity
| Query Type | OpenClaw | Hermes |
|---|
| Simple (<20 chars, e.g. "hey", "yo") | 19.1s | 34.3s (raw) / ~15s (cleaned) |
| Complex (≥20 chars) | 67.7s | 19.8s |
| With tool usage | 31.3s (1 sample) | 18.4s |
| Without tool usage | 53.2s | 25.3s |
The "Yo" Mystery Explained: Hermes' raw 34s average for simple queries is misleading. The 218s "hey" on Apr 4 at 2:58 AM was because the Qwen model wasn't loaded in LM Studio — Hermes retried 3 times, got "Model is unloaded," waited for a session reset, and only then responded. Similarly, the 303s "stop" was a context crash, not actual thinking time. With model preloaded, Hermes responds to "Good morning" in 2.1 seconds.
OpenClaw's Hidden Speed Tax: Why is OpenClaw's mean latency 52s? Context compaction. OpenClaw compacted context 6 times in 2 days, and each compaction adds 10–91 seconds of dead time. Two of those compactions took 91+ seconds — that's a minute and a half of the model just re-summarizing its own conversation. OpenClaw wasn't thinking slowly; it was doing constant housekeeping that bloated every interaction.
Suppandi was supposed to be "concise." Let's see how that went.
9.2x
verbosity ratio
User typed 1,480 chars
Suppandi responded with 13,632 chars
VS
18.8x
verbosity ratio
User typed 2,757 chars
Suppandi responded with 51,694 chars
The "Concise Servant" Report Card: Vivek asked for a "witty and sharp servant who gives concise, short responses." OpenClaw's Suppandi responds with 9x more text than the user. Hermes' Suppandi? Nearly 19x. That's not a servant — that's a politician at a town hall.
Response Length Distribution
| Metric | OpenClaw | Hermes |
|---|
| Average response length | 248 chars | 352 chars |
| Median response length | 212 chars | 110 chars |
| Maximum response length | 870 chars | 1,943 chars |
The Median Tells a Different Story: Hermes' median (110 chars) is actually shorter than OpenClaw's (212 chars) — meaning most Hermes responses are snappier. But when Hermes goes long, it goes really long (nearly 2,000 chars). OpenClaw is more consistent in length.
Running a 4B model in 6GB VRAM is like hosting a party in a studio apartment
Context Window Chaos
| Event | OpenClaw | Hermes |
|---|
| Session resets | 6 | 23 |
| Resets per day | 3.0 | 4.6 |
| Context at 100% warnings | 0 | 29 |
| Context EXCEEDED (crash) | 0 | 10 |
| "Cannot compress further" | 0 | 10 |
| Context compactions | 6 | 0 |
| Network errors | 4 | 0 |
| Retry attempts | 0 | 6 |
The Numbers Lie — OpenClaw's "Zero Crashes" is Misleading: On paper, OpenClaw shows zero context overflows. But the reality is worse: OpenClaw failed silently. Vivek reports leaving many chats where OpenClaw showed no error message — it just stopped responding. The 4 "LLM request failed: network connection error" entries on Mar 30 (where Vivek kept asking "You there?" with no response) are just the visible tip. OpenClaw's compaction was also a trap: those 6 compactions in 2 days added 10–91 seconds of dead overhead each time, making the whole experience feel sluggish and undependable.
Hermes Fails Loudly — and That's Better: Yes, Hermes hit 100% context 29 times and crashed 10 times. But it told you every single time with clear status bars and error messages. Hermes' 16K context misdetection (21 out of 24 sessions) is a real bug — it was throttling itself to 67% of available context — but even with this self-imposed handicap, its context crash recovery was fast and transparent. Hermes fails spectacularly; OpenClaw fails invisibly. For a daily-driver assistant, loud failure beats silent failure every time.
Hermes Edit/Streaming Times: Hermes messages that were edited (indicating streaming completion) took an average of 102.4 seconds to fully stream, with the worst taking 4.5 minutes. That's the full generation time for a single response on a 4B model with 6GB VRAM.
Same model, wildly different tool ecosystems
Tool Usage Comparison
| Tool Category | OpenClaw | Hermes |
|---|
| 💻 Terminal commands | 0 | 26 |
| 🌐 Browser navigation | 0 | 7 |
| 📁 File operations | 0 | 5 |
| 🧠 Memory operations | 1 | 7 |
| 📚 Skill views | 0 | 14 |
| 🐍 Code execution | 0 | 6 |
| 🔍 Session search | 0 | 2 |
Hermes is the Swiss Army Knife: Hermes used 67 tool calls across 7 categories. OpenClaw used exactly 1 tool call total (a single memory save). This isn't just a difference — it's a completely different paradigm. Hermes treats the model as an orchestrator; OpenClaw treats it as a conversationalist.
The Screenshot Saga: Vivek asked for a screenshot of his website 5 times across both agents. Neither ever delivered an actual screenshot. OpenClaw honestly said "I can't do that." Hermes tried browsers, vision tools, snapshots — burned through context — then described the website in 2,000 characters of text. Vivek: "Why vision tool? Why can't you just take a screenshot and send the file over?" 😅
When your AI servant doesn't know its own name
Identity Confusion Scorecard
| Event | OpenClaw | Hermes |
|---|
| Times user had to re-teach identity | 2 | 0 (auto-loaded) |
| Called user by wrong name | 3 ("Ganes" not "Vivek") | 0 |
| "WHO AM I?" moments | 3 | 0 |
| Failed to read identity file | 1 ("Edit failed") | 0 |
OpenClaw's Identity Crisis: Session 1: "Hey Viv! 👋 WHO AM I? And who are YOU? 🤔" — Vivek painstakingly defines Suppandi's personality. Session 2: OpenClaw reads the identity file but calls him "Ganes" instead of "Vivek" because the reasoning block decided "Ganes (shorter/friendlier)" was better. Vivek: "Why aren't you using your identity file? It has your name and my name 😭"
Hermes' Advantage: Hermes auto-loaded identity from memory and config, never needing to ask who Suppandi was. But Hermes had its own problems — on the very first day (Apr 1), it responded: "Hey! Actually, I don't have a traditional 'bootstrap sequence' where you define my personality" — completely unaware of its own framework's features.
The things that happen when you run a 4B model as your AI butler
🏆 The API Key Broadcast (Hermes): Suppandi printed the full AgentMail API key in chat 5 separate times, including in a skills list visible to anyone. The "sharp servant with security awareness" was essentially shouting credentials into Discord.
🏆 The Amavasya Time Machine (OpenClaw): Asked "When are the next 3 Amavasyas?" on March 29th, OpenClaw confidently answered: "1. January 18, 2026." That was two months in the past. Vivek: "So when are the next 3 again? You told me 2 past dates 😂." The model then searched a calendar, found the right dates, and apologized with 🙏.
🏆 "sudo deluser root" (OpenClaw): When asked about headless Ubuntu remote desktop, OpenClaw's very first command in the solution was sudo deluser root :0. Vivek: "Sudo deluser? Are you serious?" — Deleting the root user from a display group on a remote headless setup could brick the system. The "security-aware" servant nearly locked Vivek out of his own machine.
🏆 The Cron Job Talking to the Void (Hermes): Hermes set up a cron job to check AgentMail every 30 minutes. It dutifully ran 9 times. Every single response ended with: "Note: The agent cannot see this message, and therefore cannot respond to it." Several of those responses included tool calls that would never execute, and one asked for permission to set up credentials that no one would ever grant. It was an AI shouting into the void every half hour.
🏆 The Self-Repeating Loop (Hermes): In one 1,752-character message about AgentMail setup, Hermes wrote "Let me" 4 times and "Based on" 2 times, each time restarting the same explanation from scratch. It was like watching someone forget they already told you a story — mid-story — and start over. Four times.
🏆 The "Drink Water" Email (Hermes): Vivek sent a test email to Suppandi's AgentMail inbox: "Another test. Tell Vivek to drink water, Suppandi!" The cron job found it, reported it... and never told Vivek to drink water. The one instruction it received via email, and it just filed the report. 💧
🏆 The Dosa Hallucination (Hermes): Asked about nearby dosa places from Third Wave Coffee at Palladium Mall, Chennai, Hermes confidently recommended "KCR (Kannada Coffee Restaurant)" and "Dosa Kanda" — neither of which appear to be real restaurants. It then tried to run a Google Maps API call with 'YOUR_API_KEY' as the actual key. The request was flagged as dangerous and required approval. Vivek approved it. It returned: "The provided API key is invalid." Shocking.
A quantified measure of Vivek's patience
Frustration Signals Detected
| Signal | Count | Context |
|---|
| 😂 (laughing at the AI) | 2 | Past Amavasya dates, identity file |
| 😭 (crying at the AI) | 1 | "Why aren't you using your identity file?" |
| "Are you serious?" | 3 | sudo deluser, other moments |
| Bare "?" (waiting impatiently) | 1 | After long silence |
| "stop" (killing a runaway) | 1 | Hermes spiraling |
| "continue" (resuming after crash) | 1 | After context exceeded |
| "Ok? Waiting" (timeout) | 1 | Screenshot task taking too long |
| Repeated same message | 7 | Had to ask 2-3x for same thing |
Repeat Offenders: "Give me inbox contents of agentmail account" was repeated 3 times. "Get the location of the computer" was repeated 3 times. "Yo" / "hey" were repeated when the bot was unresponsive. On average, Vivek had to repeat himself once every ~11 messages. That's a 9% "did you hear me?" rate.
Who should be Suppandi's runtime?
Category Winners
| Category | Winner | Why |
|---|
| Average latency (cleaned) | Hermes | ~3x faster in practice (18.8s vs 52.2s) |
| Failure transparency | Hermes | Always tells you what's wrong; OpenClaw dies silently |
| Tool ecosystem | Hermes | 67 tool calls vs 1. Not even close. |
| Conciseness (median) | Hermes | 110 chars median vs 212 chars |
| Identity persistence | Hermes | Auto-loads identity; OpenClaw forgets |
| Context efficiency | OpenClaw | Never crashed context, but at huge latency cost |
| Overall reliability | Hermes | Crashes are visible & recoverable; OpenClaw's silent failures waste more time |
| Comedy value | Hermes | Cron jobs to the void, API key broadcasts, dosa hallucinations |
🏆 The Verdict: Hermes is a Joy
OpenClaw looks stable on paper but is a frustrating daily driver. Its "zero crash" record is misleading — it fails silently, leaving Vivek staring at Discord wondering if anything is happening. Its constant context compaction adds 10–91 seconds of dead time per cycle, making everything feel sluggish. And its single tool call (one memory save in two days) means it can't actually do anything beyond chat.
Hermes is messy but productive. It crashes, it hits context walls, it sometimes repeats itself in loops — but it always tells you what's happening, recovers quickly, and most importantly: it actually does things. Terminal commands, browser navigation, cron jobs, email integration, file management. Even when Hermes fails (and it does, loudly), the failure is actionable. You know what broke and why.
The real insight: With a 4B model on consumer hardware, the agent framework matters more than you'd think. The model is the same — it's the runtime that determines whether you get a useful assistant or a polite paperweight. OpenClaw optimizes for not crashing; Hermes optimizes for getting things done. In practice, "getting things done with occasional loud failures" beats "never crashing but never really working either."