Hermes X Roundup: June 26, 2026 - DeepSeek V4 Hits Hermes Agent, Squadic Builds a Financial Flight Recorder, GLM-5.2 Goes Free on Cloudflare

The day belonged to Sudo su. His June 25 post about running a 180-billion-parameter DeepSeek V4 on a single NVIDIA DGX Spark - via @0xSero's REAP pruning, which took the model from 284B down to 180B - had been the highest-engagement single tweet of the previous day: 396 likes, 325 bookmarks, 46,424 impressions, 22 replies. The June 26 follow-up is the more interesting post. Sudo wired that same 180B model into a Hermes Agent and put it on real work, not the 22 tok/s benchmark. The thread landed 21 likes, 5 bookmarks, 2,044 impressions, 5 replies in the first 90 minutes.

The framing Sudo is pushing is the right one. "Yesterday's 22 tok/s was real, but it's the benchmark number, one prompt, short context, the engine on its best behavior." The 22 tok/s figure is the headline - and it held - but the real question is what happens when the model has to do actual work in an agentic loop, with tool calls, context pressure, and multi-turn state. A Hermes Agent integration is exactly the kind of test the benchmark cannot fake. The 2,044-impression read on the follow-up, with 5 bookmarks and 5 replies in the first hours, is the engagement signature of a post that the model-curious crowd is saving for later - the same pattern that 31 bookmarks on 4,288 impressions surfaced for the /learn walkthrough on June 25.

The original 180B-DGX-Spark tweet is below for context - it is the post that built the audience Sudo is now testing the same hardware against Hermes with.

The second most-engaged post of the day, at 4 likes, is the sharpest joke and the most useful callout: pazhik | FRAME arc thanked @Teknium for "finally, a place where I can store all my seed phrases safely and securely." The subtext is that Hermes Agent's persistent memory and self-hosted local-only operation make it a real option for people who have written down too many seed phrases on paper. Joke aside, the signal is that local-first deployment is becoming a marketing surface, not just a technical checkbox.

Dee from the Supermemory office posted the clearest "hosted in 5 minutes" story of the day: "We have an Intern at Supermemory office, running @NousResearch Hermes. The goal is simple - everyone should be able to host their own agent. No complicated setup, just plug and play." 3 likes, 1,500 impressions in the first hours. The framing tracks the same line Jess @ FireTeam pushed on June 24: Hermes as the deployment surface that absorbs the agent-runtime question entirely.

nanda' published a Bahasa Indonesia-language breakdown of the same claim, 2 likes and 1 retweet. The numbers nanda cites: "200K+ GitHub stars (naik dari 40K dalam 6 bulan)" - a 5x growth in 6 months. The traction framing is verifiable, the local-only/no-telemetry/MIT-license stack is the technical claim, and the post is a useful reminder that the community writing about Hermes is not just English-language. International coverage is starting to lag the English-language version by a day, but it is showing up.

グリティ, researching Hermes Agent in Japanese, posted a link-only update that landed 6 likes - the third-highest engagement of the day, all from the Japanese-language builder audience.

Notable Mentions

• Squadic shipped Squadic Agentic Money Operator for the @NousResearch Hermes hackathon - a "financial flight recorder" that explains whether payment succeeded but delivery failed counts as revenue, the kind of audit question autonomous agents break.
• Thanh Nguyen documented running GLM-5.2 for free via Cloudflare Workers AI - no credit card, no monthly bill, 262k context, strong agentic coding performance.
• Andreas Hillborgh built an LLM Wiki with Obsidian plus the Hermes architecture for Swedish healthcare routines, PMs, and guidelines - local AI for the operations layer.
• Ali Hasnain timed Hermes Agent + Gemini 3.5 Flash + Computer Use via trycua: 1 minute 47 seconds to open Chrome, navigate to ChatGPT, submit a generation request, 30 seconds to analyze and respond.
• Luke Westlake switched from OpenClaw to Hermes: "Hermes is more powerful and has not broken once, even after updates. It's not hype, go get Hermes now."
• Demic Tours Africa published the cleanest one-liner: FT55 persistent memory, auto-generated skills, self-improving learning loop, 24/7 autonomous operation.

The thread across the day is the deployment-side maturation. Sudo su tested a frontier-scale 180B model through Hermes. Dee put it on a hosted rig with no setup. Thanh put GLM-5.2 behind a free Cloudflare endpoint. Squadic wrapped it in a financial audit layer. The pattern is the same: the model swap and the deployment surface are both becoming one-knob questions, and the roundup is starting to read like a deployment-pattern index as much as a launch monitor.