← Chat Archive

The `hermes update` command was blocked — the system requires your explicit approval to run it. You can run it yourself directly in your terminal: hermes update That will fetch the latest version and apply the update. If you're on a git-installed version, it'll pull from the repo and reinstall. If you'd prefer I run it on your behalf, just say "go ahead" or "approve" and I'll retry.

Update complete. Here's what happened: - 56 new commits pulled from upstream - Local changes were stashed before the pull and restored on top — no work lost - 18 stale __pycache__ directories cleaned - Python and Node.js dependencies updated - Web UI rebuilt - Bundled skills synced (1 user-modified skill kept as-is) - Config checked — already up to date One thing to note: your local changes were restored on top of the updated codebase. If anything acts weird, check `git diff` and `git status` in the hermes-agent repo directory. If the gateway was running, you may want to restart it: hermes gateway restart

Let me check the current config, test the API endpoint, and see what's going on.

Now let me test the API latency and check for known issues in parallel.

Web search hit a billing snag — let me try alternate sources and do a deeper latency test with an actual chat call.

The API endpoint is fast (54ms connect), but the web search/extract tools are blocked due to billing. Let me dig deeper — the actual chat test got HTTP 000, which may point to an auth issue.

The API endpoint is fast (54ms) but I notice the `deepseek-v4-pro` model might not be in ollama.com's model list. Let me check deeper.

The model exists on ollama.com, but the actual chat API call timed out at 30 seconds — that's the bottleneck. Let me isolate whether it's a specific model issue or a broader API problem.

Interesting! `gemma3:4b` and `deepseek-v4-flash` are fast (~0.7-0.9s), `deepseek-v3.2` is acceptable (~1.6s). But `deepseek-v4-pro` timed out at 30s. Let me verify if it's consistently failing.

Here's what I found: **The API is working** — no outages, no auth issues. The problem is the model itself. Performance comparison from direct API tests: | Model | TTFB | |---|---| | deepseek-v4-pro (current) | 3.5s+ (first attempt even timed out at 30s) | | deepseek-v4-flash | 0.7s | | deepseek-v3.2 | 1.6s | | gemma3:4b (reference) | 0.9s | deepseek-v4-pro is a massive reasoning model. Even with `reasoning: off` in your config, it still burns time generating internal reasoning tokens before it outputs anything visible. Across a multi-turn agent interaction (LLM sees tool results, decides next step, repeat), that 3-4s per turn compounds fast — a 5-tool-call task easily takes 20+ seconds just in model latency. The gateway logs also show two signal-initiated shutdowns recently (14:21 and 14:26) but no actual API errors. **Additional issue**: your Nous subscription Firecrawl/web tools are out of funds (negative balance), so web_search and web_extract are failing silently. That's separate from the LLM latency though. The fix for the slow response is to switch models. `deepseek-v4-flash` at 0.7s is functionally 5x faster per turn.

Done. Switched to `deepseek-v4-flash` and restarted the gateway. Here's the summary: - API itself is fine — no DeepSeek or ollama.com outages - The latency was entirely due to `deepseek-v4-pro` being a heavy reasoning model (3.5s+ per turn TTFB vs 0.7s for flash) - The new model is ~5x faster per turn. Over multi-tool interactions that compounds significantly - You should notice much snappier responses now Side note: your Nous subscription Firecrawl/web tool balance is negative, so web_search and web_extract fail silently with billing errors. That's unrelated to the LLM latency but worth addressing if you rely on those tools.

I'm running on **deepseek-v4-flash** via Ollama Cloud's API, which is the model we just switched to. If you want to double-check or switch again: - `hermes model` — interactive picker - `hermes config` — shows current model/provider

Big improvement. API latency dropped from 3.5s+ (and occasional 30s timeouts) on v4-pro to **~1.1s consistently** on v4-flash. Three consecutive tests all landed around 1.0-1.3s. You should feel it immediately in conversation — responses that took half a minute before will now come in a few seconds. Give it a try with a few commands and see how it feels now.