But since we have the Godot game engine at our disposal, we can do some more interesting things, like add an FPS counter, and then subsequently also let people set their preferred FPS (the idea being the potential lower power draw if someone's running it on a laptop on battery power vs someone running it on a desktop with high/native FPS). In its current state, with me using Oh-my-Pi a lot, it's evolving into a terminal workspace that can be used for orchestrating autonomous AI agents by way of dogfooding (or you know, just run herdr inside of gPTY - it's the better orchestrator and just good software - I found it after starting this project, and now I'm finding myself using it a lot).
Also, we're not limited to just terminals. Since we have Godot, we have basically a 2D (and potentially a 3D) canvas to play with. We can already full-screen the app for "zen" mode, no taskbar, no distractions. TUI die-hards can have their media or other apps entirely in terminal panes.
There has been some ground-work on getting Markdowns displayed properly done and I want to work on some kind of Wiki framework for local knowledge-management next, then create more types of panes (think native audio/video on a media pane, that sits alongside your terminal pane), and some simple 2D games (like snake) to prototype. More details are on the ROADMAP.
What's not easy (and probably won't happen) is a browser. Having done a couple of (small) projects using Electron already, the temptation to ditch Godot/Rust (learning curve) did come up (and also the ecosystem, the ease with which I could pull components and use web technologies - development velocity would definitely be higher there). But on the flipside, given all of the available LLM and AI support that we are privileged to have today, I figured the velocity should be comparable depending on how much I leaned on those. And lean I did.
Godot/Rust seemed the better call to me and my intent anyway - going with the 'it's not just the end but the journey that matters' philosophy. So yes, there has been heavy use of LLMs & AI to generate a lot of the code. But I do review and steer actively, not relying solely on vibes, and there's a few bits here & there that have been human authored.
There are definitely a lot of polish and QoL items that need to land to make the end user experience better, but in the meantime, let me know your thoughts and/or concerns!
Repository: https://github.com/godot-pty/gpty
Docs/Blog: https://godot-pty.github.io/gpty/"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Godot and Rust based multiplexer (terminal panes and more)"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/godot-pty/gpty"}},"_tags":["story","author_1nv1n","story_49660676","show_hn"],"author":"1nv1n","children":[49660677,49660948,49661044,49661214,49661263,49661307,49661327,49661705,49661710,49662081,49662113,49662190,49662725,49662845,49662882,49663145,49664340,49665949,49667526,49669020,49669446,49680099],"created_at":"2026-09-11T16:03:05Z","created_at_i":1789142585,"num_comments":50,"objectID":"49660676","points":97,"story_id":49660676,"story_text":"I wanted to share my side-project: gPTY. Started off as an idea to combine Godot and Rust in a project (two stacks I wanted to use more to learn more). The base inspiration was tmux - simply allow spawning multiple PTYs and then let the user grid/tile them how they see fit.
But since we have the Godot game engine at our disposal, we can do some more interesting things, like add an FPS counter, and then subsequently also let people set their preferred FPS (the idea being the potential lower power draw if someone's running it on a laptop on battery power vs someone running it on a desktop with high/native FPS). In its current state, with me using Oh-my-Pi a lot, it's evolving into a terminal workspace that can be used for orchestrating autonomous AI agents by way of dogfooding (or you know, just run herdr inside of gPTY - it's the better orchestrator and just good software - I found it after starting this project, and now I'm finding myself using it a lot).
Also, we're not limited to just terminals. Since we have Godot, we have basically a 2D (and potentially a 3D) canvas to play with. We can already full-screen the app for "zen" mode, no taskbar, no distractions. TUI die-hards can have their media or other apps entirely in terminal panes.
There has been some ground-work on getting Markdowns displayed properly done and I want to work on some kind of Wiki framework for local knowledge-management next, then create more types of panes (think native audio/video on a media pane, that sits alongside your terminal pane), and some simple 2D games (like snake) to prototype. More details are on the ROADMAP.
What's not easy (and probably won't happen) is a browser. Having done a couple of (small) projects using Electron already, the temptation to ditch Godot/Rust (learning curve) did come up (and also the ecosystem, the ease with which I could pull components and use web technologies - development velocity would definitely be higher there). But on the flipside, given all of the available LLM and AI support that we are privileged to have today, I figured the velocity should be comparable depending on how much I leaned on those. And lean I did.
Godot/Rust seemed the better call to me and my intent anyway - going with the 'it's not just the end but the journey that matters' philosophy. So yes, there has been heavy use of LLMs & AI to generate a lot of the code. But I do review and steer actively, not relying solely on vibes, and there's a few bits here & there that have been human authored.
There are definitely a lot of polish and QoL items that need to land to make the end user experience better, but in the meantime, let me know your thoughts and/or concerns!
Repository: https://github.com/godot-pty/gpty
Docs/Blog: https://godot-pty.github.io/gpty/","title":"Show HN: Godot and Rust based multiplexer (terminal panes and more)","updated_at":"2026-09-21T17:05:00Z","url":"https://github.com/godot-pty/gpty"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"jatora"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["oh","my","pi"],"value":"I run a lot of coding agents at once, mostly Claude Code, Codex and pi, across several projects. I got tired of having so many windows open, of agents leaving hung processes behind, of tracking which one needed input, and of having to remote into my desktop or SSH into a single shell to manage them from my phone. Even then, a terminal on a phone couldn't show me the frontend or output I needed to verify the work. I tried Orca and herdr. Both were better than my setup, but neither had a real WYSIWYG markdown notes surface, and neither had a phone interface I could actually work in.
swe-mux is the terminal multiplexer I built for that. It runs any shell or CLI in real terminals on your own machine and serves the same live sessions to your desktop and your phone. It's built around agents: Claude Code, Codex, opencode, pi and oh-my-pi get a live status (working, done, or waiting on you), searchable history and resume. Anything else runs as an ordinary terminal with everything except the status.
This is the actual terminal and not a UI wrapped around the agent, and you can just use it that way, or start claude in a plain shell and that pane becomes an agent session.
On the desktop it's panes and tabs of sessions, notes, files and dev-server previews. Input behaves the same across agents: Shift+Enter, Ctrl+Backspace, paste and click-to-position work in all of them, shortcuts come from presets (tmux, VS Code, Vim, Emacs), and you can paste images straight from the clipboard.
Notes and any markdown file in the project open in the same WYSIWYG editor, on desktop and phone, and autosave. No more scattered locations for notes.
On the phone it's the same sessions over your own Tailscale network, installed as a PWA: the terminal with a usable keyboard, git diff review, files, notes, and push notifications when an agent needs a decision. A dev server on the host's 127.0.0.1 is proxied through swe-mux, so the phone can open it without exposing another port. Being able to play-test and check a frontend there is what finally closed the development loop on the phone for a lot of my projects. Now I can and do swe anywhere. It's a problem
Two things I use constantly: a prompt queue, so something I think of mid-turn waits until the agent is done instead of interrupting it, and messaging between sessions, so a Codex session can ask a Claude session something. I don't want my workflow tied to one provider.
General architecture:
Sessions survive restarts. A separate supervisor process owns the terminals, so the daemon can restart and the app can update without killing anything. I update it several times a day, so this wasn't optional.
Status comes from the CLI's hooks, its transcript and the terminal itself, because no single one reliably tells "waiting on you" apart from "busy in the background".
There's no swe-mux account, no server of mine in the path, and no telemetry. Just your tailnet
It was built on Windows, which is where I use it every day and the only platform with a packaged desktop app. The installer is unsigned, so the PyPI install is the way to avoid the SmartScreen prompt. macOS and Linux run the daemon and use a browser. CI runs it on all three, but I haven't lived on those myself, however Linux I have tested.
Demo, the real UI with simulated agents, nothing to install: https://swemux.dev/demo/
Install: `uv tool install swe-mux` (or `pipx install swe-mux`), then run `swe-mux` on Windows or `swemux start` elsewhere.
Apache-2.0: https://github.com/jatoran/swe-mux
I've put a lot of work into the initial user experience to help streamline a quick setup depending on someone's needs, feedback greatly appreciated, thanks"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: swe-mux \u2013 A terminal multiplexer for coding agents optimized for mobile"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/jatoran/swe-mux"}},"_tags":["story","author_jatora","story_49867905","show_hn"],"author":"jatora","created_at":"2026-09-27T16:02:04Z","created_at_i":1790524924,"num_comments":0,"objectID":"49867905","points":3,"story_id":49867905,"story_text":"I run a lot of coding agents at once, mostly Claude Code, Codex and pi, across several projects. I got tired of having so many windows open, of agents leaving hung processes behind, of tracking which one needed input, and of having to remote into my desktop or SSH into a single shell to manage them from my phone. Even then, a terminal on a phone couldn't show me the frontend or output I needed to verify the work. I tried Orca and herdr. Both were better than my setup, but neither had a real WYSIWYG markdown notes surface, and neither had a phone interface I could actually work in.
swe-mux is the terminal multiplexer I built for that. It runs any shell or CLI in real terminals on your own machine and serves the same live sessions to your desktop and your phone. It's built around agents: Claude Code, Codex, opencode, pi and oh-my-pi get a live status (working, done, or waiting on you), searchable history and resume. Anything else runs as an ordinary terminal with everything except the status.
This is the actual terminal and not a UI wrapped around the agent, and you can just use it that way, or start claude in a plain shell and that pane becomes an agent session.
On the desktop it's panes and tabs of sessions, notes, files and dev-server previews. Input behaves the same across agents: Shift+Enter, Ctrl+Backspace, paste and click-to-position work in all of them, shortcuts come from presets (tmux, VS Code, Vim, Emacs), and you can paste images straight from the clipboard.
Notes and any markdown file in the project open in the same WYSIWYG editor, on desktop and phone, and autosave. No more scattered locations for notes.
On the phone it's the same sessions over your own Tailscale network, installed as a PWA: the terminal with a usable keyboard, git diff review, files, notes, and push notifications when an agent needs a decision. A dev server on the host's 127.0.0.1 is proxied through swe-mux, so the phone can open it without exposing another port. Being able to play-test and check a frontend there is what finally closed the development loop on the phone for a lot of my projects. Now I can and do swe anywhere. It's a problem
Two things I use constantly: a prompt queue, so something I think of mid-turn waits until the agent is done instead of interrupting it, and messaging between sessions, so a Codex session can ask a Claude session something. I don't want my workflow tied to one provider.
General architecture:
Sessions survive restarts. A separate supervisor process owns the terminals, so the daemon can restart and the app can update without killing anything. I update it several times a day, so this wasn't optional.
Status comes from the CLI's hooks, its transcript and the terminal itself, because no single one reliably tells "waiting on you" apart from "busy in the background".
There's no swe-mux account, no server of mine in the path, and no telemetry. Just your tailnet
It was built on Windows, which is where I use it every day and the only platform with a packaged desktop app. The installer is unsigned, so the PyPI install is the way to avoid the SmartScreen prompt. macOS and Linux run the daemon and use a browser. CI runs it on all three, but I haven't lived on those myself, however Linux I have tested.
Demo, the real UI with simulated agents, nothing to install: https://swemux.dev/demo/
Install: `uv tool install swe-mux` (or `pipx install swe-mux`), then run `swe-mux` on Windows or `swemux start` elsewhere.
Apache-2.0: https://github.com/jatoran/swe-mux
I've put a lot of work into the initial user experience to help streamline a quick setup depending on someone's needs, feedback greatly appreciated, thanks","title":"Show HN: swe-mux \u2013 A terminal multiplexer for coding agents optimized for mobile","updated_at":"2026-09-27T23:35:48Z","url":"https://github.com/jatoran/swe-mux"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"opwizardx"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["oh","my","pi"],"value":"I wanted all of my sessions preserved in one S3 storage dump I could access from anywhere. I couldn't find anything that could store sessions from multiple machines in one remote S3 bucket without running a database service anywhere, make them available as an MCP for my agent, and efficiently search through them. pond was born out of this, and out of my necessity to stop being locked to my laptop. I'm convinced that the sessions we generate with agents are one of the most precious assets we have and it still feels so odd to me that we throw them away like we couldn't care less.
With pond you just point it at an S3 bucket or a local directory. Under the hood is in-process Lance. It also supports safe concurrent writes, so I just collect all my sessions from all my local machines, remote VMs and telegram agent setups in one bucket dir. Right now that is 14,861 sessions, 2.83M messages, 10.6 GiB, from 8 tools (Claude Code, Codex, opencode, pi, Claude Desktop, oh-my-pi and a few more).
The one thing I really want to get optimized is remote storage read speeds. Local queries run in milliseconds to two seconds on my whole corpus, while remote storage is still 20-30 seconds per call. My goal is to get them on par.
What I would really want to know is which client to support next, and whether keeping the vector search and all of that embedded-model hassle is worth its keep. FTS-only would have been much faster, and there is hardly any good benchmark I can find that could help me close this question."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Pond \u2013 lossless archive for agent sessions in your own S3"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/tenequm/pond"}},"_tags":["story","author_opwizardx","story_49376500","show_hn"],"author":"opwizardx","children":[49391256],"created_at":"2026-08-20T16:04:21Z","created_at_i":1787241861,"num_comments":0,"objectID":"49376500","points":3,"story_id":49376500,"story_text":"I wanted all of my sessions preserved in one S3 storage dump I could access from anywhere. I couldn't find anything that could store sessions from multiple machines in one remote S3 bucket without running a database service anywhere, make them available as an MCP for my agent, and efficiently search through them. pond was born out of this, and out of my necessity to stop being locked to my laptop. I'm convinced that the sessions we generate with agents are one of the most precious assets we have and it still feels so odd to me that we throw them away like we couldn't care less.
With pond you just point it at an S3 bucket or a local directory. Under the hood is in-process Lance. It also supports safe concurrent writes, so I just collect all my sessions from all my local machines, remote VMs and telegram agent setups in one bucket dir. Right now that is 14,861 sessions, 2.83M messages, 10.6 GiB, from 8 tools (Claude Code, Codex, opencode, pi, Claude Desktop, oh-my-pi and a few more).
The one thing I really want to get optimized is remote storage read speeds. Local queries run in milliseconds to two seconds on my whole corpus, while remote storage is still 20-30 seconds per call. My goal is to get them on par.
What I would really want to know is which client to support next, and whether keeping the vector search and all of that embedded-model hassle is worth its keep. FTS-only would have been much faster, and there is hardly any good benchmark I can find that could help me close this question.","title":"Show HN: Pond \u2013 lossless archive for agent sessions in your own S3","updated_at":"2026-08-21T17:24:48Z","url":"https://github.com/tenequm/pond"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"mrsirg"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["oh","my","pi"],"value":"yeah everyone is building there own harness. it\u2019s fun though, owning the whole stack. tried opencode, pi, oh my pi, prime agent. wrote some extensions for pi. just wasn\u2019t doing it for me. so I built my own agent runtime. and now it\u2019s turned into my daily driver. Been dogfooding it for a month now and it\u2019s the only harness I want to use. zero bloat, concurrent tool calls, plugins. latest release avail for download (only for macOS/linux).
https://mrsirg97-rgb.github.io/rig/\nhttps://github.com/mrsirg97-rgb/rig"},"title":{"matchLevel":"none","matchedWords":[],"value":"Building your own agent harness is fun"}},"_tags":["story","author_mrsirg","story_49667385","ask_hn"],"author":"mrsirg","children":[49667392,49667712],"created_at":"2026-09-12T00:41:30Z","created_at_i":1789173690,"num_comments":2,"objectID":"49667385","points":2,"story_id":49667385,"story_text":"yeah everyone is building there own harness. it\u2019s fun though, owning the whole stack. tried opencode, pi, oh my pi, prime agent. wrote some extensions for pi. just wasn\u2019t doing it for me. so I built my own agent runtime. and now it\u2019s turned into my daily driver. Been dogfooding it for a month now and it\u2019s the only harness I want to use. zero bloat, concurrent tool calls, plugins. latest release avail for download (only for macOS/linux).
https://mrsirg97-rgb.github.io/rig/\nhttps://github.com/mrsirg97-rgb/rig","title":"Building your own agent harness is fun","updated_at":"2026-09-12T02:30:30Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"arush15june"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["oh","my","pi"],"value":"I am 4.1 maxxing on commandcode GOAT Plan + api rates with oh my pi for the last 4 weeks, it's absolutely amazing and crazy fast, it's alright if it makes a mistake, I have enough time to iterate again, I have also added an advisor layer of mimo 2.6 pro which does make it a notch smarter. Getting haiku 5.5/sonnet5.5 to work on plans and letting 4.1 flash work through it is helping a ton too.
I am a big ChatGPT fan, all our team has ChatGPT Subs, but the TPS across all models including luna is just so damn slow.
Commandcode giving 60$ worth of Deepseek for 10$ is just genuinely goat.
And it never says no for cyber tasks so that's a big win"},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Why isn't the industry freaking out about DeepSeek 4.1 Flash?"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://www.dgt.is/blog/2026-10-07-deepseek-freek-out/"}},"_tags":["comment","author_arush15june","story_50000488"],"author":"arush15june","comment_text":"I am 4.1 maxxing on commandcode GOAT Plan + api rates with oh my pi for the last 4 weeks, it's absolutely amazing and crazy fast, it's alright if it makes a mistake, I have enough time to iterate again, I have also added an advisor layer of mimo 2.6 pro which does make it a notch smarter. Getting haiku 5.5/sonnet5.5 to work on plans and letting 4.1 flash work through it is helping a ton too.
I am a big ChatGPT fan, all our team has ChatGPT Subs, but the TPS across all models including luna is just so damn slow.
Commandcode giving 60$ worth of Deepseek for 10$ is just genuinely goat.
And it never says no for cyber tasks so that's a big win","created_at":"2026-10-08T21:14:02Z","created_at_i":1791494042,"objectID":"50012393","parent_id":50000488,"story_id":50000488,"story_title":"Why isn't the industry freaking out about DeepSeek 4.1 Flash?","story_url":"https://www.dgt.is/blog/2026-10-07-deepseek-freek-out/","updated_at":"2026-10-08T21:30:56Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"jeremyjh"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["oh","my","pi"],"value":"I use oh-my-pi (omp.sh) - mostly with stock settings and skills but its a "fully loaded" harness that you don't really have to add anything to. I do change a couple of things: I set it to prefer subagents, and I enable rewind. It is crazy that rewind is not a default, it saves a TON of context - if the agent goes down a crazy path that burns a lot of tokens it can rewind to an earlier checkpoint with the exploration or bug hunting summary.
Presently I'm using sol-high for the default agent which does orchestration and a lot of smaller investigation and coding tasks itself. sol-max for planning and review. Luna-max for planned coding and general tasks. I also have a $10 minimax plan and use M3 for exploration and library roles, but I could probably be using Luna for that just as well and still only very rarely run into usage issues.
I don't use any plugins or skill libraries apart from Caveman and I'm not sure how useful that really is anymore so I'd start without it so you have a baseline to compare. I do think it reduces context usage a bit but I haven't measured it recently. Caveman also includes some team, agent & investigation skills - again they might be helping but I haven't re-evaluated since like 90 days ago."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Claude Haiku 5.5"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://www.anthropic.com/claude-haiku-5-5"}},"_tags":["comment","author_jeremyjh","story_49996437"],"author":"jeremyjh","comment_text":"I use oh-my-pi (omp.sh) - mostly with stock settings and skills but its a "fully loaded" harness that you don't really have to add anything to. I do change a couple of things: I set it to prefer subagents, and I enable rewind. It is crazy that rewind is not a default, it saves a TON of context - if the agent goes down a crazy path that burns a lot of tokens it can rewind to an earlier checkpoint with the exploration or bug hunting summary.
Presently I'm using sol-high for the default agent which does orchestration and a lot of smaller investigation and coding tasks itself. sol-max for planning and review. Luna-max for planned coding and general tasks. I also have a $10 minimax plan and use M3 for exploration and library roles, but I could probably be using Luna for that just as well and still only very rarely run into usage issues.
I don't use any plugins or skill libraries apart from Caveman and I'm not sure how useful that really is anymore so I'd start without it so you have a baseline to compare. I do think it reduces context usage a bit but I haven't measured it recently. Caveman also includes some team, agent & investigation skills - again they might be helping but I haven't re-evaluated since like 90 days ago.","created_at":"2026-10-08T11:18:22Z","created_at_i":1791458302,"objectID":"50004447","parent_id":50001004,"story_id":49996437,"story_title":"Claude Haiku 5.5","story_url":"https://www.anthropic.com/claude-haiku-5-5","updated_at":"2026-10-08T11:23:25Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"haellsigh"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["oh","my","pi"],"value":"It looks like he uses oh-my-pi, which is a full featured version of pi and has this to-do style out of the box."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/Niko1221/Strata"}},"_tags":["comment","author_haellsigh","story_49953495"],"author":"haellsigh","comment_text":"It looks like he uses oh-my-pi, which is a full featured version of pi and has this to-do style out of the box.","created_at":"2026-10-05T07:01:48Z","created_at_i":1791183708,"objectID":"49961552","parent_id":49961324,"story_id":49953495,"story_title":"Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s","story_url":"https://github.com/Niko1221/Strata","updated_at":"2026-10-05T07:03:27Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"Winfred-zz"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["oh","my","pi"],"value":"I just ran a set of benchmarks, ninfer-3090-qwen3.8-27b (so mix of Q4 and Q5) vs strata-qwen3.8-flash-next-iq3_xxs (so Q3):
\u2502------------------- \u2502 Ninfer-3090 \u2502 Strata
\u2502 Code generation \u2502 52/78 (66.7%) \u2502 70/78 (89.7%)
\u2502 Code completion \u2502 40/50 (80.0%) \u2502 44/50 (88.0%)
\u2502 Total------------- \u2502 92/128 (71.9%) \u2502 114/128 (89.1%)
\u2502 API failures------ \u2502 10 \u2502 5
- Ninfer generation: ~122 min total.
- Strata generation: ~142 min total.
So strata is a little slower, but keep in mind that ninfer-3090 is very optimized for a Qwen 3.8. Standard Qwen 3.8 runs at 20 t/s, this modified version can do 50 t/s (but it's extremely long in it's thinking, it just goes on and on.
This is on a 3090 that will crash unless power capped, with a Zen 2 CPU, 64GB DDR4 with a PCIe that refuses to go higher than 8x (basically pretty crappy all in all).
Yet with some tweaking and optimizing I still manage to get strata to run at 40 to 60 t/s.
That strata has been optimized on my Oh My Pi conversations. So when I'm using it, it's probably faster and closer to ninfer in speed than during those unoptimized benchmark tests."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/Niko1221/Strata"}},"_tags":["comment","author_Winfred-zz","story_49953495"],"author":"Winfred-zz","children":[49958674,49959827,49960046,49961097,49961154],"comment_text":"I just ran a set of benchmarks, ninfer-3090-qwen3.8-27b (so mix of Q4 and Q5) vs strata-qwen3.8-flash-next-iq3_xxs (so Q3):
\u2502------------------- \u2502 Ninfer-3090 \u2502 Strata
\u2502 Code generation \u2502 52/78 (66.7%) \u2502 70/78 (89.7%)
\u2502 Code completion \u2502 40/50 (80.0%) \u2502 44/50 (88.0%)
\u2502 Total------------- \u2502 92/128 (71.9%) \u2502 114/128 (89.1%)
\u2502 API failures------ \u2502 10 \u2502 5
- Ninfer generation: ~122 min total.
- Strata generation: ~142 min total.
So strata is a little slower, but keep in mind that ninfer-3090 is very optimized for a Qwen 3.8. Standard Qwen 3.8 runs at 20 t/s, this modified version can do 50 t/s (but it's extremely long in it's thinking, it just goes on and on.
This is on a 3090 that will crash unless power capped, with a Zen 2 CPU, 64GB DDR4 with a PCIe that refuses to go higher than 8x (basically pretty crappy all in all).
Yet with some tweaking and optimizing I still manage to get strata to run at 40 to 60 t/s.
That strata has been optimized on my Oh My Pi conversations. So when I'm using it, it's probably faster and closer to ninfer in speed than during those unoptimized benchmark tests.","created_at":"2026-10-04T21:34:00Z","created_at_i":1791149640,"objectID":"49958120","parent_id":49955565,"story_id":49953495,"story_title":"Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s","story_url":"https://github.com/Niko1221/Strata","updated_at":"2026-10-07T02:26:32Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"zahrevsky"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["oh","my","pi"],"value":"There is a command in oh-my-pi called "/omfg <problem>". You explain what is wrong with agent's response, and it writes a hook to make sure that the problem doesn't happen again. It then re-runs your previous prompt to make sure that hook is triggered, and if not, it rewrites the hook to make your previous prompt trigger the hook. Then each next agent's response is checked by the hook, and if it is triggered, the agent receives feedback on what's wrong and what must be done differently."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Agents don't need memory, they need documentation"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://liao.gg/blog/agents-dont-need-memory"}},"_tags":["comment","author_zahrevsky","story_49945933"],"author":"zahrevsky","children":[49950577,49951134,49952632],"comment_text":"There is a command in oh-my-pi called "/omfg <problem>". You explain what is wrong with agent's response, and it writes a hook to make sure that the problem doesn't happen again. It then re-runs your previous prompt to make sure that hook is triggered, and if not, it rewrites the hook to make your previous prompt trigger the hook. Then each next agent's response is checked by the hook, and if it is triggered, the agent receives feedback on what's wrong and what must be done differently.","created_at":"2026-10-04T00:49:46Z","created_at_i":1791074986,"objectID":"49949426","parent_id":49949330,"story_id":49945933,"story_title":"Agents don't need memory, they need documentation","story_url":"https://liao.gg/blog/agents-dont-need-memory","updated_at":"2026-10-08T12:33:38Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"ttoinou"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["oh","my","pi"],"value":"Ive been using this since it was initially released with deepseek v4 flash, and it is absolutely the best launcher ever on my m5 max 128gb
Now Ive been running qwen 3.8 flash next for more than a week and it\u2019s doing great, really fast and super long context windows. Sometimes the model is behaving stupidly by not remembering something I said earlier but it could be also a problem from the agentic AI harness. Im using oh my pi but Im wondering what people are using ds4 with here ?"},"story_title":{"matchLevel":"none","matchedWords":[],"value":"From the creator of Redis; run LLM locally with ds4"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://dwarfstar.sh/"}},"_tags":["comment","author_ttoinou","story_49936575"],"author":"ttoinou","children":[49941769],"comment_text":"Ive been using this since it was initially released with deepseek v4 flash, and it is absolutely the best launcher ever on my m5 max 128gb
Now Ive been running qwen 3.8 flash next for more than a week and it\u2019s doing great, really fast and super long context windows. Sometimes the model is behaving stupidly by not remembering something I said earlier but it could be also a problem from the agentic AI harness. Im using oh my pi but Im wondering what people are using ds4 with here ?","created_at":"2026-10-02T22:19:03Z","created_at_i":1790979543,"objectID":"49939328","parent_id":49936575,"story_id":49936575,"story_title":"From the creator of Redis; run LLM locally with ds4","story_url":"https://dwarfstar.sh/","updated_at":"2026-10-05T07:15:12Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"dom96"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["oh","my","pi"],"value":"I\u2019ve tried Pi and honestly it felt pretty rough and raw. Kind of like Arch Linux. Even with oh-my-pi the experience just didn\u2019t seem polished enough for my taste. I\u2019m not really sure why I would use it over OpenCode."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Pi 1.0"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://earendil.com/posts/pi-1-0/"}},"_tags":["comment","author_dom96","story_49926069"],"author":"dom96","children":[49932745],"comment_text":"I\u2019ve tried Pi and honestly it felt pretty rough and raw. Kind of like Arch Linux. Even with oh-my-pi the experience just didn\u2019t seem polished enough for my taste. I\u2019m not really sure why I would use it over OpenCode.","created_at":"2026-10-02T12:17:35Z","created_at_i":1790943455,"objectID":"49932649","parent_id":49926069,"story_id":49926069,"story_title":"Pi 1.0","story_url":"https://earendil.com/posts/pi-1-0/","updated_at":"2026-10-02T15:33:18Z"}],"hitsPerPage":20,"nbHits":236,"nbPages":12,"page":0,"params":"query=oh-my-pi&advancedSyntax=true&analyticsTags=backend","processingTimeMS":26,"processingTimingsMS":{"_request":{"roundTrip":23},"afterFetch":{"format":{"highlighting":1,"total":1}},"fetch":{"scanning":24,"total":25},"total":26},"query":"oh-my-pi","serverTimeMS":28}