Built my own Qwen3.8-27B engine for my RTX 3090: 100+ tok/s decode at 200k context, 1.2-1.3x faster than HyperQwen's mtp-long at long context (same card, single stream). Now working on better 4-bit KV quality.
Check it out: tuan3w/qwen-warp
1 like · 49 views
Last weekend I let AI agents run on one question: can autonomous research build a faster local Qwen server than vLLM?
~2 days later, on one RTX 3090: 184 vs 97 tok/s on code (up to ~2x).
Code + full worklog (experiments, dead ends): tuan3w/qwen-warp
1 like · 40 views
@davidfano I had bookmarked tuan3w/awesome-tech-rss for the Tech theme, but even that is only scratching the surface.
And then you'd need the same principle applied to any theme of interest.
Lots to do!
1 like · 12 views
herdr mobile: GPL-3.0 Android app to supervise your coding agents over plain SSH
I wanted one app on my phone for the coding agents on my own machines, with no cloud, no account and no relay. So I built herdr mobile.
- Every agent on every machine in one list, with whatever needs you at the top
- Answer permission requests from the phone, always with the exact command. Risky ones (git push, rm, "always allow") need a hold
- Start and chat with agents over ACP: omp, Claude Code, Codex (omp is the most tested)
- Live terminal view, SFTP file browser, photos and files from the phone
- Plain SSH, one channel per machine, works well over Tailscale
Repo and APK: tuan3w/herdr-mobile
Needs herdr ( herdr.dev ) on the host. Android only for now. Feedback and PRs welcome.
herdr mobile: GPL-3.0 Android app to supervise your coding agents over plain SSH
preview.redd.it
I wanted one app on my phone for the coding agents on my own machines, with no cloud, no account and no relay. So I built herdr mobile.
- Every agent on every machine in one list, with whatever needs you at the top
- Answer permission requests from the phone, always with the exact command. Risky ones (git push, rm, "always allow") need a hold
- Start and chat with agents over ACP: omp, Claude Code, Codex (omp is the most tested)
- Live terminal view, SFTP file browser, photos and files from the phone
- Plain SSH, one channel per machine, works well over Tailscale
Demo on my phone (video is made by hyperframes): youtube.com
Repo and APK: tuan3w/herdr-mobile
Needs herdr (herdr.dev) on the host. Android only for now. Feedback and PRs welcome.
Mentions
Built my own Qwen3.8-27B engine for my RTX 3090: 100+ tok/s decode at 200k context, 1.2-1.3x faster than HyperQwen's mtp-long at long context (same card, single stream). Now working on better 4-bit KV quality. Check it out: tuan3w/qwen-warp
Last weekend I let AI agents run on one question: can autonomous research build a faster local Qwen server than vLLM? ~2 days later, on one RTX 3090: 184 vs 97 tok/s on code (up to ~2x). Code + full worklog (experiments, dead ends): tuan3w/qwen-warp
@davidfano I had bookmarked tuan3w/awesome-tech-rss for the Tech theme, but even that is only scratching the surface. And then you'd need the same principle applied to any theme of interest. Lots to do!
herdr mobile: GPL-3.0 Android app to supervise your coding agents over plain SSH I wanted one app on my phone for the coding agents on my own machines, with no cloud, no account and no relay. So I built herdr mobile. - Every agent on every machine in one list, with whatever needs you at the top - Answer permission requests from the phone, always with the exact command. Risky ones (git push, rm, "always allow") need a hold - Start and chat with agents over ACP: omp, Claude Code, Codex (omp is the most tested) - Live terminal view, SFTP file browser, photos and files from the phone - Plain SSH, one channel per machine, works well over Tailscale Repo and APK: tuan3w/herdr-mobile Needs herdr ( herdr.dev ) on the host. Android only for now. Feedback and PRs welcome.
herdr mobile: GPL-3.0 Android app to supervise your coding agents over plain SSH preview.redd.it I wanted one app on my phone for the coding agents on my own machines, with no cloud, no account and no relay. So I built herdr mobile. - Every agent on every machine in one list, with whatever needs you at the top - Answer permission requests from the phone, always with the exact command. Risky ones (git push, rm, "always allow") need a hold - Start and chat with agents over ACP: omp, Claude Code, Codex (omp is the most tested) - Live terminal view, SFTP file browser, photos and files from the phone - Plain SSH, one channel per machine, works well over Tailscale Demo on my phone (video is made by hyperframes): youtube.com Repo and APK: tuan3w/herdr-mobile Needs herdr (herdr.dev) on the host. Android only for now. Feedback and PRs welcome.