I skipped a week, so this issue carries two. the short version: ten new wingtips, a release that closed every open P0 and P1, Hermes Cloud going live, Hermes Agent turning up on the rabbit r1, and four nights of training a speculative-decode draft head on one RTX 5090. the local-ai corner is at the bottom as usual.

this fortnight's tips

ten Hermes Wingtips shipped across the two weeks, each one tested on my own boxes before it went out:

• #9: a profile is a separate agent: one profile with many gateways is still one agent everywhere; a profile per platform splits it into separate agents, and --clone-all forks everything across.
• #10: fallback chains skip half-filled entries: every entry needs both provider and model or Hermes Agent skips it; put a local model last as the floor that never rate-limits.
• #11: the kanban toolset is opt-in: toolsets: all doesn't include it, name kanban explicitly to let an agent drive the board.
• #12: kanban worker workspaces are scratch: wiped on completion by design; keep output with --workspace dir:<absolute path>.
• #13: the session db grows on purpose: auto-prune ships off so history stays searchable; hermes sessions prune trims ended sessions when it gets heavy.
• #14: switching models resets your prompt cache: the cache key includes the model, so the next turn re-reads the whole conversation at full input price; batch your switches on long threads.
• #15: what a compaction actually keeps: it rewrites history into a structured working brief (goal, constraints, progress, decisions, next steps) while the raw turns stay in state.db.
• #16: gateway gone after a reboot: a user-service gateway runs under your login and stops with it; sudo loginctl enable-linger $USER keeps it up.
• #17: check the built-in tools before installing a skill: memory, web search, browser control, cron and sub-agents are already native.
• #18: keep auto-learning, review its writes: /memory approval on and /skills approval on stage every write for your yes.

shipped this fortnight

• Hermes Logbook #1: the first issue of my between-release changelog digest, covering what merged since v0.17.0. the release notes lag the merges, the Logbook doesn't.
• Hermes v0.18.0, "The Judgment Release": every P0 and P1 closed, 692 issues cleared, 998 PRs from 370+ contributors.
• faster, cheaper web reading: web_extract now truncates and stores full content instead of re-fetching (PR #54843).
• the NVIDIA DGX Spark playbook: an A-to-Z guide for running Hermes with local vLLM on the DGX Spark.

around the ecosystem

• Hermes Cloud is live: hosted Hermes, one login, no box to babysit. my rewrite of the launch and why it matters if you've been putting off the self-host step.
• rabbitOS 2.3 ships Hermes Agent on the r1: your Hermes Agent in your pocket, on hardware you already own.

local AI beyond Hermes

the 5090 had a busy fortnight. the headline: I'm training the first speculative-decode draft head for Hermes-4.3-36B, on the same card that serves it.

• EAGLE-3 draft head, night 1: an EAGLE-3 draft head trains on the teacher's internal activations, and caching them to disk for the standard 54K-sample recipe wants 4.5TB. I had 570GB free. so: online mode, the 4-bit teacher (19.5GiB, SGLang) stays resident and feeds the draft live, while the fp32 optimiser state (8.4GiB) sits in system RAM with the step on CPU. night 1 gave 7,000 steps in 6h50m, expected acceptance 0.05 to 0.27 and climbing. bonus find: FSDP allocates gradient storage for frozen params, 1.5GiB for an embedding that never updates, so single-GPU skips the wrap entirely.
• the nightly grind: night 2 doubled the training context to 2048, night 3 passed 21,000 of 54,000 steps, and on night 4, halfway through the epoch, the curve split.
• the sm_120 field guide: 6 weeks of running everything AI on one RTX 5090, and every wall I hit is now a public reference.
• one RTX 5090 generates a full 4-minute song in 1.75s.
• video generation with synchronised sound now runs on a gaming GPU.
• Z-Image-Turbo on the RTX 5090 (sm_120).
• DeepSeek-OCR-2, tried locally.
• my Qwopus-3.6-35B-A3B-Coder-MTP benchmark numbers.

back to the weekly cadence from here. if one of these tips saved you an evening, forward this to the friend still debugging their gateway at midnight.