This started as an exploration into harnesses and locally hostable LLMs (the kind that uses less power than a small electric heater), and I used it as an opportunity to retool. I've tried popular ones like Cline, Claude Code, and Pi, but wanted a different UX and integrated sandboxes.
The solution that I've landed on is a responsive web interface where I can quickly drive progress. It feels like a microcosm of my day job: define and delegate tasks, organize using kanban, then review deliverables. And it also feels like the time I used to spend coding is now spent in code review with the agent, so I tried to optimize the UX to make review feel nicer.
One of my (maybe a bit niche) benchmark use-cases is how pleasant it is to make progress on hobby software projects while also out on a hike. This meant specifying low bandwidth usage and offline mode.
Now I can jot down an idea, start the task, and at some point down the trail: get a notification* that either clarification is needed, a decision needs to be made, or the code is ready for review. I've been having a good time with this.
To build MIKE, I exclusively used locally hostable LLMs (from Google, Meta, Qwen, Poolside, and DeepSeek) hosted with llama.cpp on my machines at home (Strix Halo and a 5090) to "compile" the code. The experience has gotten substantially better ever since DeepSeek\u2019s release in July. Side note: I have a soft spot for llama.cpp since it was one of the first that made it possible to run this new tech at home, and its maintainers accepted a tiny improvement I made to error reporting. Oh the joys of having a PR get approved.
I plan on maintaining this, at least until similar features are supported by better tools. Until then, hope you find this useful!
5 minute demo: https://youtu.be/jexzLkZ50FU
p.s.
I didn't mention details about the sandbox earlier since most people seem to find this a bit boring. It was one of my critical requirements, and I opted to use an existing battle-tested sandbox. I'm not an expert on how the Linux kernel works, but seeing all the VM escape exploits recently is a bit scary, and I liked gVisor\u2019s approach to the problem.
I'm also using Git worktrees to provision the agent\u2019s workspace, as well as a read-only repository mount to provide relevant context when the agent looks for it.
*off by default"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["agent"],"value":"Show HN: Cosplay as a tech lead, using gVisor to sandbox agents"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/SoftwareRenderer/MIKE"}},"_tags":["story","author_SftwareRenderer","story_50137043","show_hn"],"author":"SftwareRenderer","children":[50137381],"created_at":"2026-10-18T01:49:34Z","created_at_i":1792288174,"num_comments":1,"objectID":"50137043","points":2,"story_id":50137043,"story_text":"Hi everyone,
This started as an exploration into harnesses and locally hostable LLMs (the kind that uses less power than a small electric heater), and I used it as an opportunity to retool. I've tried popular ones like Cline, Claude Code, and Pi, but wanted a different UX and integrated sandboxes.
The solution that I've landed on is a responsive web interface where I can quickly drive progress. It feels like a microcosm of my day job: define and delegate tasks, organize using kanban, then review deliverables. And it also feels like the time I used to spend coding is now spent in code review with the agent, so I tried to optimize the UX to make review feel nicer.
One of my (maybe a bit niche) benchmark use-cases is how pleasant it is to make progress on hobby software projects while also out on a hike. This meant specifying low bandwidth usage and offline mode.
Now I can jot down an idea, start the task, and at some point down the trail: get a notification* that either clarification is needed, a decision needs to be made, or the code is ready for review. I've been having a good time with this.
To build MIKE, I exclusively used locally hostable LLMs (from Google, Meta, Qwen, Poolside, and DeepSeek) hosted with llama.cpp on my machines at home (Strix Halo and a 5090) to "compile" the code. The experience has gotten substantially better ever since DeepSeek\u2019s release in July. Side note: I have a soft spot for llama.cpp since it was one of the first that made it possible to run this new tech at home, and its maintainers accepted a tiny improvement I made to error reporting. Oh the joys of having a PR get approved.
I plan on maintaining this, at least until similar features are supported by better tools. Until then, hope you find this useful!
5 minute demo: https://youtu.be/jexzLkZ50FU
p.s.
I didn't mention details about the sandbox earlier since most people seem to find this a bit boring. It was one of my critical requirements, and I opted to use an existing battle-tested sandbox. I'm not an expert on how the Linux kernel works, but seeing all the VM escape exploits recently is a bit scary, and I liked gVisor\u2019s approach to the problem.
I'm also using Git worktrees to provision the agent\u2019s workspace, as well as a read-only repository mount to provide relevant context when the agent looks for it.
*off by default","title":"Show HN: Cosplay as a tech lead, using gVisor to sandbox agents","updated_at":"2026-10-18T02:46:08Z","url":"https://github.com/SoftwareRenderer/MIKE"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"hephaes7us"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["qwen","coding","agent"],"value":"I consider "agentic" to simply mean that it performs well given a loop of actions and a success criterion. As such, lots of code _can_ be written agentically (even if it is not the best or fastest way to do it).
When I am wrapping up work for the day, I usually set Qwen3.8-27B loose on some coding task. It's very good about continuing its agent loop until it can plausibly claim success. A lot of this work I might end up throwing away, but pretty often there's some piece of it that's worth keeping. Because it was "agentic", I end up with some useful pieces for little effort.
A less agentic model would need me involved the whole time, and while my being there would probably steer it towards initially better results, I'm not working overnight anyway. So, autonomous/agentic but not-so-smart and not-always-correct is still valuable to me.
Anyway, I've found that it's reasonably easy to constrain a "highly agentic" model to suit interactive usage, but back when they were less agentic it was annoying to get them to be autonomous. So I think more agentic is better, provided the instruction-adherence is strong enough."},"story_title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["agent"],"value":"LFM2.5-24B-A2B: Built for fast on-device agents"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://www.liquid.ai/blog/lfm2-5-24b-a2b"}},"_tags":["comment","author_hephaes7us","story_50097369"],"author":"hephaes7us","children":[50104997],"comment_text":"I consider "agentic" to simply mean that it performs well given a loop of actions and a success criterion. As such, lots of code _can_ be written agentically (even if it is not the best or fastest way to do it).
When I am wrapping up work for the day, I usually set Qwen3.8-27B loose on some coding task. It's very good about continuing its agent loop until it can plausibly claim success. A lot of this work I might end up throwing away, but pretty often there's some piece of it that's worth keeping. Because it was "agentic", I end up with some useful pieces for little effort.
A less agentic model would need me involved the whole time, and while my being there would probably steer it towards initially better results, I'm not working overnight anyway. So, autonomous/agentic but not-so-smart and not-always-correct is still valuable to me.
Anyway, I've found that it's reasonably easy to constrain a "highly agentic" model to suit interactive usage, but back when they were less agentic it was annoying to get them to be autonomous. So I think more agentic is better, provided the instruction-adherence is strong enough.","created_at":"2026-10-15T15:09:31Z","created_at_i":1792076971,"objectID":"50104826","parent_id":50104282,"story_id":50097369,"story_title":"LFM2.5-24B-A2B: Built for fast on-device agents","story_url":"https://www.liquid.ai/blog/lfm2-5-24b-a2b","updated_at":"2026-10-15T15:22:32Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"JCKodel"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["qwen","coding","agent"],"value":"Hi HN, I'm the author. FACE is a single Go binary. You run `face` in\nthe folder that holds your projects and it serves a page on\n127.0.0.1. It speaks ACP (the Agent Client Protocol that Zed\npublished) to the official CLI of the agent you already use: Claude\nCode, Codex, Copilot, Cursor, OpenCode, Qwen Code or Antigravity,\nsigned in with your own account. FACE has no model, no key and no\nagent loop of its own.
Why I built it: a coding agent ships a feature in an afternoon, and\nthree people and five sessions later nobody can answer "what is in\nflight, what did it cost, why was it decided this way". The answer I\nlanded on is not a dashboard but a method where every decision and\nevery delivery is a versioned file: a handful of docs written once, a\nqueue with one line per delivery (docs/06), and one page per delivery\n(work/<slug>.md) that says what will be built and how it is proven\nbefore any code changes. The method is four slash commands the agent\nreads from one file (brainstorm, analyze, propose, apply); it works\nwithout FACE. FACE is the window: the queue becomes a five-lane board,\na card opens into the page plus the live session, and the commit\nwindow shows exactly what the agent staged. The agent never commits\nor merges; you click.
I built FACE itself with it, from the fourth week on. Before that the\nsame method shipped a job marketplace in 89 one-page deliveries over\n14 days, and a consulting project in 107 deliveries over 19 days.
Install: `brew install jckodel/face/face`, `winget install\nJCKodel.FACE`, a `.deb`, or `curl -fsSL\nhttps://face.kodel.com.br/install.sh | sh`. macOS, Linux, Windows,\narm64 and amd64. UI in en, pt, es, fr, de and zh-Hans.
What is missing: single user and local only (a team mode with a\nserver on your network is the next milestone); no editor (it shows\ncode and diffs, your editor stays your editor); the source is not\npublic yet. It will be visible under a fair source license when the\npaid tier for teams exists, and the free tier stays free. By itself\nFACE makes at most two requests, both off until you turn them on: a\ndaily anonymous ping that counts machines and a daily version check.\nThe page "What FACE sends" lists both URLs.
Happy to answer anything about ACP, the adapters for Claude and Codex\n(they are not ACP native, so FACE wraps them), or the method."},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["agent"],"value":"Show HN: Face, a browser window over agents that keeps the plan in files"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://face.kodel.com.br"}},"_tags":["story","author_JCKodel","story_50102683","show_hn"],"author":"JCKodel","created_at":"2026-10-15T12:29:09Z","created_at_i":1792067349,"num_comments":0,"objectID":"50102683","points":1,"story_id":50102683,"story_text":"Hi HN, I'm the author. FACE is a single Go binary. You run `face` in\nthe folder that holds your projects and it serves a page on\n127.0.0.1. It speaks ACP (the Agent Client Protocol that Zed\npublished) to the official CLI of the agent you already use: Claude\nCode, Codex, Copilot, Cursor, OpenCode, Qwen Code or Antigravity,\nsigned in with your own account. FACE has no model, no key and no\nagent loop of its own.
Why I built it: a coding agent ships a feature in an afternoon, and\nthree people and five sessions later nobody can answer "what is in\nflight, what did it cost, why was it decided this way". The answer I\nlanded on is not a dashboard but a method where every decision and\nevery delivery is a versioned file: a handful of docs written once, a\nqueue with one line per delivery (docs/06), and one page per delivery\n(work/<slug>.md) that says what will be built and how it is proven\nbefore any code changes. The method is four slash commands the agent\nreads from one file (brainstorm, analyze, propose, apply); it works\nwithout FACE. FACE is the window: the queue becomes a five-lane board,\na card opens into the page plus the live session, and the commit\nwindow shows exactly what the agent staged. The agent never commits\nor merges; you click.
I built FACE itself with it, from the fourth week on. Before that the\nsame method shipped a job marketplace in 89 one-page deliveries over\n14 days, and a consulting project in 107 deliveries over 19 days.
Install: `brew install jckodel/face/face`, `winget install\nJCKodel.FACE`, a `.deb`, or `curl -fsSL\nhttps://face.kodel.com.br/install.sh | sh`. macOS, Linux, Windows,\narm64 and amd64. UI in en, pt, es, fr, de and zh-Hans.
What is missing: single user and local only (a team mode with a\nserver on your network is the next milestone); no editor (it shows\ncode and diffs, your editor stays your editor); the source is not\npublic yet. It will be visible under a fair source license when the\npaid tier for teams exists, and the free tier stays free. By itself\nFACE makes at most two requests, both off until you turn them on: a\ndaily anonymous ping that counts machines and a daily version check.\nThe page "What FACE sends" lists both URLs.
Happy to answer anything about ACP, the adapters for Claude and Codex\n(they are not ACP native, so FACE wraps them), or the method.","title":"Show HN: Face, a browser window over agents that keeps the plan in files","updated_at":"2026-10-15T12:31:30Z","url":"https://face.kodel.com.br"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"eyegor"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["qwen","coding","agent"],"value":"Don't waste time on bonsai unless you have like 16gb or less ram (ram, not vram). Although I'm pretty sure you could still run this model well on 16+8, and qwen3.6 35b-a3b runs well on ~32gb combined ram at q4. Just offload most of the experts to cpu/ram if you have limited vram budget. With the moe models it doesn't hurt performance too much.
But bonsai is basically trash on real agentic coding work. I've tried both bonsai models, all the quants, messed with the hyperparameters, tried to fine tune (didn't succeed), and it just kept looping or failing to finish tasks. Qwen3.6 35b-a3b is my current go to for agentic work for any systems under 24gb vram, although honestly I'm not sure the dense qwen3.8 27b is meaningfully better. The main difference I see is it does a lot more testing attempts and burns a ton of tokens + compactions, so every task it takes 1.2-2x as long as 35b-a3b does. But the quality of the results is comparable."},"story_title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["agent"],"value":"LFM2.5-24B-A2B: Built for fast on-device agents"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://www.liquid.ai/blog/lfm2-5-24b-a2b"}},"_tags":["comment","author_eyegor","story_50097369"],"author":"eyegor","children":[50102220,50101749,50103073],"comment_text":"Don't waste time on bonsai unless you have like 16gb or less ram (ram, not vram). Although I'm pretty sure you could still run this model well on 16+8, and qwen3.6 35b-a3b runs well on ~32gb combined ram at q4. Just offload most of the experts to cpu/ram if you have limited vram budget. With the moe models it doesn't hurt performance too much.
But bonsai is basically trash on real agentic coding work. I've tried both bonsai models, all the quants, messed with the hyperparameters, tried to fine tune (didn't succeed), and it just kept looping or failing to finish tasks. Qwen3.6 35b-a3b is my current go to for agentic work for any systems under 24gb vram, although honestly I'm not sure the dense qwen3.8 27b is meaningfully better. The main difference I see is it does a lot more testing attempts and burns a ton of tokens + compactions, so every task it takes 1.2-2x as long as 35b-a3b does. But the quality of the results is comparable.","created_at":"2026-10-15T05:20:32Z","created_at_i":1792041632,"objectID":"50099500","parent_id":50098829,"story_id":50097369,"story_title":"LFM2.5-24B-A2B: Built for fast on-device agents","story_url":"https://www.liquid.ai/blog/lfm2-5-24b-a2b","updated_at":"2026-10-16T00:18:47Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"davidsainez"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["qwen","coding","agent"],"value":"AFAIU prefill speed makes a bigger difference in agentic coding workflows as the harness likely spends more time reading than creating new code. Qwen in particular spends a lot of time reasoning so it does make a larger impact there"},"story_title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["agent"],"value":"LFM2.5-24B-A2B: Built for fast on-device agents"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://www.liquid.ai/blog/lfm2-5-24b-a2b"}},"_tags":["comment","author_davidsainez","story_50097369"],"author":"davidsainez","comment_text":"AFAIU prefill speed makes a bigger difference in agentic coding workflows as the harness likely spends more time reading than creating new code. Qwen in particular spends a lot of time reasoning so it does make a larger impact there","created_at":"2026-10-15T01:14:14Z","created_at_i":1792026854,"objectID":"50097912","parent_id":50097719,"story_id":50097369,"story_title":"LFM2.5-24B-A2B: Built for fast on-device agents","story_url":"https://www.liquid.ai/blog/lfm2-5-24b-a2b","updated_at":"2026-10-15T01:20:00Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"AspireOne"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["qwen","coding","agent"],"value":"Yes, that's the reason. Anthropic said as much in their statement (they said that there's a tension between writing human-understandable prose and prose optimal for other agents to read).
As a side note, I am kind of sad that basically all models nowadays are RLd to hell and back specifically for coding and deep reasoning, and as a result, they lack in creative writing, human-readable prose, and even general knowledge and abilities (for example, qwen 3.8 27b is a beast for coding, but reportedly falls off the cliff in anything else). Gemini 4 Argon seems to be an example of what you get when you DON'T do that - Google seems to be one of the only labs optimizing for general performance.
I also fear that models cannot be excellent in everything at the same time. I fear that there will always be a heavy tradeoff between what specific field the labs optimize for VS everything else. I don't know."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Gemini 4 Argon"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://deepmind.google/models/gemini/"}},"_tags":["comment","author_AspireOne","story_50073879"],"author":"AspireOne","children":[50089064],"comment_text":"Yes, that's the reason. Anthropic said as much in their statement (they said that there's a tension between writing human-understandable prose and prose optimal for other agents to read).
As a side note, I am kind of sad that basically all models nowadays are RLd to hell and back specifically for coding and deep reasoning, and as a result, they lack in creative writing, human-readable prose, and even general knowledge and abilities (for example, qwen 3.8 27b is a beast for coding, but reportedly falls off the cliff in anything else). Gemini 4 Argon seems to be an example of what you get when you DON'T do that - Google seems to be one of the only labs optimizing for general performance.
I also fear that models cannot be excellent in everything at the same time. I fear that there will always be a heavy tradeoff between what specific field the labs optimize for VS everything else. I don't know.","created_at":"2026-10-14T14:53:01Z","created_at_i":1791989581,"objectID":"50088604","parent_id":50082853,"story_id":50073879,"story_title":"Gemini 4 Argon","story_url":"https://deepmind.google/models/gemini/","updated_at":"2026-10-15T10:00:45Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"Xeoncross"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["qwen","coding","agent"],"value":"At this point, I really want all AI benchmarks to use the best coding/writing model that normal people can run locally.
For now that seems to be Qwen 3.8 27B that requires ~28GB of RAM (Q4) which is pretty doable for anyone with a beefy GPU or slightly-above-average mac.
I want to know how close open weights (open models) are to all these leading commercial models. Yes, yes, local models aren't as good, but that distance is closing as local models are already better than what we had just over a year ago in many cases.
Same with Text-to-Image generation (Flux 2, Z-Image-Turbo, etc..) or text-to-speech (Parakeet, Whisper, etc..)."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Parsimony \u2013 an open-source benchmark for how much code AI models add"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://parsimony.lasu.dev/"}},"_tags":["comment","author_Xeoncross","story_50077619"],"author":"Xeoncross","children":[50081260],"comment_text":"At this point, I really want all AI benchmarks to use the best coding/writing model that normal people can run locally.
For now that seems to be Qwen 3.8 27B that requires ~28GB of RAM (Q4) which is pretty doable for anyone with a beefy GPU or slightly-above-average mac.
I want to know how close open weights (open models) are to all these leading commercial models. Yes, yes, local models aren't as good, but that distance is closing as local models are already better than what we had just over a year ago in many cases.
Same with Text-to-Image generation (Flux 2, Z-Image-Turbo, etc..) or text-to-speech (Parakeet, Whisper, etc..).","created_at":"2026-10-14T01:21:51Z","created_at_i":1791940911,"objectID":"50081234","parent_id":50077619,"story_id":50077619,"story_title":"Show HN: Parsimony \u2013 an open-source benchmark for how much code AI models add","story_url":"https://parsimony.lasu.dev/","updated_at":"2026-10-14T02:26:56Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"tempoponet"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["qwen","coding","agent"],"value":"Most popular would be Qwen 9b or Qwen 35b with experts offloaded to CPU. Even these aren't great, but they can be run at higher quants on lower hardware. We'll see what Qwen releases version 4 in the coming weeks."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Sub-1-Bit LLM Compression via Latent Factorization"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/SamsungLabs/LittleBit"}},"_tags":["comment","author_tempoponet","story_50005608"],"author":"tempoponet","children":[50020643],"comment_text":"Most popular would be Qwen 9b or Qwen 35b with experts offloaded to CPU. Even these aren't great, but they can be run at higher quants on lower hardware. We'll see what Qwen releases version 4 in the coming weeks.","created_at":"2026-10-08T20:48:54Z","created_at_i":1791492534,"objectID":"50012051","parent_id":50010530,"story_id":50005608,"story_title":"Sub-1-Bit LLM Compression via Latent Factorization","story_url":"https://github.com/SamsungLabs/LittleBit","updated_at":"2026-10-09T13:58:13Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"netsroht"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["qwen","coding","agent"],"value":"I'm getting about 3k tok/s on prefill so it heavily depends on the input size. I hooked it up into a coding agent for shell permission checks, estimated shell execution time (buckets), goal pre-screening, subagent routing etc. and I have RTTs between 100-200ms. It can take noticeably longer for huge contexts because of the prefill speed but it can be cached so subsequent calls will be much faster.
I tried smaller decision models like laya (also with custom finetunes) but the accuracy was not really good (for the things I tested). Also I don't have any VRAM left on this GPU so i had to decide whether to host laya or qwen-3.8-27b but not both at the same time. Running decision models on a CPU will also be noticeably slower so I went down this route to have both combined with shared base weights."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Decisions API is in public beta"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://developers.openai.com/api/docs/guides/decisions"}},"_tags":["comment","author_netsroht","story_49984025"],"author":"netsroht","comment_text":"I'm getting about 3k tok/s on prefill so it heavily depends on the input size. I hooked it up into a coding agent for shell permission checks, estimated shell execution time (buckets), goal pre-screening, subagent routing etc. and I have RTTs between 100-200ms. It can take noticeably longer for huge contexts because of the prefill speed but it can be cached so subsequent calls will be much faster.
I tried smaller decision models like laya (also with custom finetunes) but the accuracy was not really good (for the things I tested). Also I don't have any VRAM left on this GPU so i had to decide whether to host laya or qwen-3.8-27b but not both at the same time. Running decision models on a CPU will also be noticeably slower so I went down this route to have both combined with shared base weights.","created_at":"2026-10-07T09:37:00Z","created_at_i":1791365820,"objectID":"49990367","parent_id":49990196,"story_id":49984025,"story_title":"Decisions API is in public beta","story_url":"https://developers.openai.com/api/docs/guides/decisions","updated_at":"2026-10-07T09:40:48Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"silentcuriousde"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["qwen","coding","agent"],"value":"I recently open-sourced my coding agent CLI that I've been working on for the past while. It is a project I never intended to make public since it was part of my own personal local AI stack. But as I chipped away at it and made it actually usable as a daily driver, I thought it would be nice to make it public for others to see and use.
I initially built it to see how much I could get the harness to make small models, especially at low quantization and context to not feel terrible to use. So while Reika doesn't solve the intelligence side (it never will), it tries to solve the overall experience when using small models at the absolute scale.
A lot of the testing and pain came through working on my M2 MacBook Air 16GB trying to run models like Qwen3.6 35B A3B and Qwen3.8 27B all day in agentic coding, maxing out the RAM and limits of my own machine. So the base of Reika comes from a legitimate source of truth.
You can also plug in an API key for those with hybrid setups too."},"story_title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["coding","agent"],"value":"Show HN: Reika \u2013 A coding agent CLI designed around small local models first"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/alexwkleung/reika"}},"_tags":["comment","author_silentcuriousde","story_49982904"],"author":"silentcuriousde","comment_text":"I recently open-sourced my coding agent CLI that I've been working on for the past while. It is a project I never intended to make public since it was part of my own personal local AI stack. But as I chipped away at it and made it actually usable as a daily driver, I thought it would be nice to make it public for others to see and use.
I initially built it to see how much I could get the harness to make small models, especially at low quantization and context to not feel terrible to use. So while Reika doesn't solve the intelligence side (it never will), it tries to solve the overall experience when using small models at the absolute scale.
A lot of the testing and pain came through working on my M2 MacBook Air 16GB trying to run models like Qwen3.6 35B A3B and Qwen3.8 27B all day in agentic coding, maxing out the RAM and limits of my own machine. So the base of Reika comes from a legitimate source of truth.
You can also plug in an API key for those with hybrid setups too.","created_at":"2026-10-06T19:29:25Z","created_at_i":1791314965,"objectID":"49982905","parent_id":49982904,"story_id":49982904,"story_title":"Show HN: Reika \u2013 A coding agent CLI designed around small local models first","story_url":"https://github.com/alexwkleung/reika","updated_at":"2026-10-06T19:31:47Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"ineptech"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["qwen","coding","agent"],"value":"I follow the same approach (finding LLMs useful but only running open models on my own hardware) but I've been limited by also not letting them into my email and being pretty restrictive about their internet access. Would you mind sharing what stack and agent harness you use, since your other comments make it clear you're highly restrictive about what kinds of tech you run at home and I assume you vetted it thoroughly, and it sounds like it's working well?
I'm wondering specifically about things like what kind of tool or restrictions you use on web browsing to avoid being flagged as a bot and how you guardrail access to your email (to avoid, say, the LLM deciding that "help me organize" includes telling people you want to reschedule). I use Qwen 3.8 26b at home and find it very useful for some tasks but when I think of using it as a general assistant rather than coding assistant, I keep thinking of times it has gone wildly wrong (e.g. it once spent 20 minutes trying to get around Supabase authentication because it couldn't figure why a certain call wasn't working)"},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Meta\u2019s Muse is an adorable privacy and security dumpster fire"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://www.techdirt.com/2026/10/06/metas-muse-is-an-adorable-privacy-and-security-dumpster-fire/"}},"_tags":["comment","author_ineptech","story_49977588"],"author":"ineptech","children":[49985305],"comment_text":"I follow the same approach (finding LLMs useful but only running open models on my own hardware) but I've been limited by also not letting them into my email and being pretty restrictive about their internet access. Would you mind sharing what stack and agent harness you use, since your other comments make it clear you're highly restrictive about what kinds of tech you run at home and I assume you vetted it thoroughly, and it sounds like it's working well?
I'm wondering specifically about things like what kind of tool or restrictions you use on web browsing to avoid being flagged as a bot and how you guardrail access to your email (to avoid, say, the LLM deciding that "help me organize" includes telling people you want to reschedule). I use Qwen 3.8 26b at home and find it very useful for some tasks but when I think of using it as a general assistant rather than coding assistant, I keep thinking of times it has gone wildly wrong (e.g. it once spent 20 minutes trying to get around Supabase authentication because it couldn't figure why a certain call wasn't working)","created_at":"2026-10-06T17:26:23Z","created_at_i":1791307583,"objectID":"49981532","parent_id":49980512,"story_id":49977588,"story_title":"Meta\u2019s Muse is an adorable privacy and security dumpster fire","story_url":"https://www.techdirt.com/2026/10/06/metas-muse-is-an-adorable-privacy-and-security-dumpster-fire/","updated_at":"2026-10-06T22:53:32Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"firelex"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["qwen","coding","agent"],"value":"For anyone interested, we tried the adapter approach in a coding agent harness: https://jeffhub.ai/notes/jeff-v1-3
What's interesting is that these are two adapters: Jeff making simple decisions (which tool to call, how to parameterize it) and Jeff routing between different version of Qwen. There are lots of applications for this approach.
In terms of Jeff-Code specifically, treat it as an experiment. Swift-Qwen is faster, but also - I understand - worse on benchmarks. The interesting bit is the concept here."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Jeff \u2013 Jev-compatible 0.8B decision models, trained at home, ~30 ms"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/firelex/jeff"}},"_tags":["comment","author_firelex","story_49883844"],"author":"firelex","comment_text":"For anyone interested, we tried the adapter approach in a coding agent harness: https://jeffhub.ai/notes/jeff-v1-3
What's interesting is that these are two adapters: Jeff making simple decisions (which tool to call, how to parameterize it) and Jeff routing between different version of Qwen. There are lots of applications for this approach.
In terms of Jeff-Code specifically, treat it as an experiment. Swift-Qwen is faster, but also - I understand - worse on benchmarks. The interesting bit is the concept here.","created_at":"2026-10-06T09:31:50Z","created_at_i":1791279110,"objectID":"49976177","parent_id":49883844,"story_id":49883844,"story_title":"Jeff \u2013 Jev-compatible 0.8B decision models, trained at home, ~30 ms","story_url":"https://github.com/firelex/jeff","updated_at":"2026-10-06T09:34:00Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"janalsncm"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["qwen","coding","agent"],"value":"The reality is they trained a model and it looks worse on benchmarks than Qwen or GLM. I don\u2019t see how sharing the weights hurts anyone? Even when Llama 4 came out and it was a dumpster fire, it didn\u2019t affect me personally.
> That only means their training regime is inferior if their predecessors did so much more with so much less
Hard to imagine how that wouldn\u2019t be the case. They probably missed the boat on distilling Claude (or their lawyers said no), they probably didn\u2019t hire an army of math PhDs to write reasoning traces, they don\u2019t have millions of DAUs in a coding agent to train from, and they probably have less money, less experience, fewer top tier researchers, and fewer resources for experiments. They are an underdog without a doubt.
None of that means they shouldn\u2019t release their model."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Beam: Reflection's 501B open-weight model"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://reflection.ai/blog/introducing-beam"}},"_tags":["comment","author_janalsncm","story_49969183"],"author":"janalsncm","children":[49973206],"comment_text":"The reality is they trained a model and it looks worse on benchmarks than Qwen or GLM. I don\u2019t see how sharing the weights hurts anyone? Even when Llama 4 came out and it was a dumpster fire, it didn\u2019t affect me personally.
> That only means their training regime is inferior if their predecessors did so much more with so much less
Hard to imagine how that wouldn\u2019t be the case. They probably missed the boat on distilling Claude (or their lawyers said no), they probably didn\u2019t hire an army of math PhDs to write reasoning traces, they don\u2019t have millions of DAUs in a coding agent to train from, and they probably have less money, less experience, fewer top tier researchers, and fewer resources for experiments. They are an underdog without a doubt.
None of that means they shouldn\u2019t release their model.","created_at":"2026-10-05T20:56:51Z","created_at_i":1791233811,"objectID":"49970605","parent_id":49970325,"story_id":49969183,"story_title":"Beam: Reflection's 501B open-weight model","story_url":"https://reflection.ai/blog/introducing-beam","updated_at":"2026-10-06T11:31:01Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"silentcuriousde"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["qwen","coding","agent"],"value":"I recently open-sourced my coding agent CLI that I've been working on for the past while. It is a project I never intended to make public since it was part of my own personal local AI stack. But as I chipped away at it and made it actually usable as a daily driver, I thought it would be nice to make it public for others to see and use.
I initially built it to see how much I could get the harness to make small models, especially at low quantization and context to not feel terrible to use. So while Reika doesn't solve the intelligence side (it never will), it tries to solve the overall experience when using small models at the absolute scale.
A lot of the testing and pain came through working on my M2 MacBook Air 16GB trying to run models like Qwen3.6 35B A3B and Qwen3.8 27B all day in agentic coding, maxing out the RAM and limits of my own machine. So the base of Reika comes from a legitimate source of truth.
For those of you that don't care about local AI, you can always plug in an API key into Reika and enjoy it that way.
And it's not perfect and I don't claim it to be something evolutionary either. It's just another option among many other harnesses out there.
Anyways, thanks for reading :)"},"story_title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["coding","agent"],"value":"Reika \u2013 A coding agent CLI designed around small local models first"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/alexwkleung/reika"}},"_tags":["comment","author_silentcuriousde","story_49968417"],"author":"silentcuriousde","comment_text":"I recently open-sourced my coding agent CLI that I've been working on for the past while. It is a project I never intended to make public since it was part of my own personal local AI stack. But as I chipped away at it and made it actually usable as a daily driver, I thought it would be nice to make it public for others to see and use.
I initially built it to see how much I could get the harness to make small models, especially at low quantization and context to not feel terrible to use. So while Reika doesn't solve the intelligence side (it never will), it tries to solve the overall experience when using small models at the absolute scale.
A lot of the testing and pain came through working on my M2 MacBook Air 16GB trying to run models like Qwen3.6 35B A3B and Qwen3.8 27B all day in agentic coding, maxing out the RAM and limits of my own machine. So the base of Reika comes from a legitimate source of truth.
For those of you that don't care about local AI, you can always plug in an API key into Reika and enjoy it that way.
And it's not perfect and I don't claim it to be something evolutionary either. It's just another option among many other harnesses out there.
Anyways, thanks for reading :)","created_at":"2026-10-05T18:16:31Z","created_at_i":1791224191,"objectID":"49968418","parent_id":49968417,"story_id":49968417,"story_title":"Reika \u2013 A coding agent CLI designed around small local models first","story_url":"https://github.com/alexwkleung/reika","updated_at":"2026-10-05T18:19:46Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"hecturchi"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["qwen","coding","agent"],"value":"- Tiny context size or hours to load it
- Hard to benefit from thinking and preserve thinking given token cost.
- Low quants reduce accuracy heavMTP draft can make it make the same mistakes all the time when calling tools, formatting output or following basic guidelines. Otherwise 2x-4x slower.
- K/V quants probably quantized too make things less accurate.
Useful would be combinations with:
- Full context size so it can code and think a bit.
- Draft MTP <= 2 so it doesn't trip
- Q4 quants or better so its accurate
- q8 cache or better so it stays accurate.
- 20 token/s so it finishes while reviewing previous step.
- 1000 tokens/s context load so compactions don't waste 10+ minutes.
- And enough left RAM for 50+ context checkpoints so that it can progress quuckly.
Closest you have is Qwen3.6-35B-A3B-MTP.
Latest gens (Qwen3.8 and co.) are just too big for low specs. 27B dense models seem to be ok for integrated >=92 GiB RAM.
Source: I have low specs and tried them all for agentic use + coding."},"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_hecturchi","story_49953495"],"author":"hecturchi","children":[49956194,49956264,49955936,49956539,49956725,49956770],"comment_text":"- Tiny context size or hours to load it
- Hard to benefit from thinking and preserve thinking given token cost.
- Low quants reduce accuracy heavMTP draft can make it make the same mistakes all the time when calling tools, formatting output or following basic guidelines. Otherwise 2x-4x slower.
- K/V quants probably quantized too make things less accurate.
Useful would be combinations with:
- Full context size so it can code and think a bit.
- Draft MTP <= 2 so it doesn't trip
- Q4 quants or better so its accurate
- q8 cache or better so it stays accurate.
- 20 token/s so it finishes while reviewing previous step.
- 1000 tokens/s context load so compactions don't waste 10+ minutes.
- And enough left RAM for 50+ context checkpoints so that it can progress quuckly.
Closest you have is Qwen3.6-35B-A3B-MTP.
Latest gens (Qwen3.8 and co.) are just too big for low specs. 27B dense models seem to be ok for integrated >=92 GiB RAM.
Source: I have low specs and tried them all for agentic use + coding.","created_at":"2026-10-04T17:27:25Z","created_at_i":1791134845,"objectID":"49955898","parent_id":49953495,"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":"2026-10-06T17:34:02Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"stymaar"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["qwen","coding","agent"],"value":"It depends on what you mean by \u201cOpus-like\u201d, because if you mean \u201cas strong as Opus 4.6 for agentic coding\u201d then Qwen3.8-27B has been there for the past two months.
But if you mean \u201cas strong as current-gen Opus\u201d then it's probably never gonna happen, but it doesn't really matter since we're long into the diminishing returns for performance improvements: I haven't notice any major leap between 4.6 and 5.5 in my daily usage, and I'm convinced that with a fact enough piecs of hardware I would be using local Qwen exclusively (I'm using it daily but only at night for long running tasks because they take much more time than Opus due to the compounding effects of my slow GPU and Qwen's verbosity)."},"story_title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["qwen"],"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_stymaar","story_49953495"],"author":"stymaar","children":[49954996,49955024,49955031],"comment_text":"It depends on what you mean by \u201cOpus-like\u201d, because if you mean \u201cas strong as Opus 4.6 for agentic coding\u201d then Qwen3.8-27B has been there for the past two months.
But if you mean \u201cas strong as current-gen Opus\u201d then it's probably never gonna happen, but it doesn't really matter since we're long into the diminishing returns for performance improvements: I haven't notice any major leap between 4.6 and 5.5 in my daily usage, and I'm convinced that with a fact enough piecs of hardware I would be using local Qwen exclusively (I'm using it daily but only at night for long running tasks because they take much more time than Opus due to the compounding effects of my slow GPU and Qwen's verbosity).","created_at":"2026-10-04T15:26:40Z","created_at_i":1791127600,"objectID":"49954714","parent_id":49954498,"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":"2026-10-06T02:43:46Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"cyanydeez"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["qwen","coding","agent"],"value":"I'm working on local models; qwen3.8-Flash-Next seems to be passed to rubicon.
I'm working on javascript, and now I'm only doing this via typescript. I've successfully got this going:
1. Design a feature in plain language with the coding agen (opencode) and write a document for it.
2. Restart the context (or I use /compact to flush any errant details)
3. Pull the new plan into context and ask the agent to revise it follow Test Driven Development.
4. Depending on how big it is, the agent places it into multi stages, each with it's own document.
It then loops through the stages. It might be entirely based on the language you're using, but this loop seems strong enough.
Then when there's bugs, errors, anything, we update the doc, add more tests, then revise the code.
It's quite possible the harness you're using isn't setup properly. For me to do this locally, I had to hack on dynamic context pruning which I described here: https://news.ycombinator.com/item?id=49906637#49907641"},"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_cyanydeez","story_49945933"],"author":"cyanydeez","comment_text":"I'm working on local models; qwen3.8-Flash-Next seems to be passed to rubicon.
I'm working on javascript, and now I'm only doing this via typescript. I've successfully got this going:
1. Design a feature in plain language with the coding agen (opencode) and write a document for it.
2. Restart the context (or I use /compact to flush any errant details)
3. Pull the new plan into context and ask the agent to revise it follow Test Driven Development.
4. Depending on how big it is, the agent places it into multi stages, each with it's own document.
It then loops through the stages. It might be entirely based on the language you're using, but this loop seems strong enough.
Then when there's bugs, errors, anything, we update the doc, add more tests, then revise the code.
It's quite possible the harness you're using isn't setup properly. For me to do this locally, I had to hack on dynamic context pruning which I described here: https://news.ycombinator.com/item?id=49906637#49907641","created_at":"2026-10-04T12:41:03Z","created_at_i":1791117663,"objectID":"49953422","parent_id":49953270,"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":"2026-10-04T12:58:24Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"stkdump"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["qwen","coding","agent"],"value":"I have a computer with a 5090 on a smart plug at home running Qwen3.8 27B for agentic coding. On a busy day it can use around 5kWh, though on most days it is around 2kWh. I am sure cloud is more efficient because there is probably efficiency in running many parallel streams, some of the models have fever than 27B active parameters and the power draw of the non-GPU components is also spread over more GPUs. But still I also believe that the energy use numbers you get from the inference providers are a bit "beautified". After all, they still fight a political battle and have to show that it isn't all so bad.
Having said that, seeing the incredible progress of models throughout this year, I also strongly believe that the planned buildout is overeager. Even I with my gaming hardware often run out of instructions to give. And the smarter the models get that I can run, the less I will be able to saturate my hardware. Is it because of my lack of creativity of which kinds of tasks I can give to AI? Maybe a bit, but currently I can't believe that I am that far away."},"story_title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["coding"],"value":"One month coding with GLM 5.3 Flash"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://wagtail.org/blog/one-month-on-glm-53-flash/"}},"_tags":["comment","author_stkdump","story_49934620"],"author":"stkdump","children":[49942161,49943766,49942442,49948626],"comment_text":"I have a computer with a 5090 on a smart plug at home running Qwen3.8 27B for agentic coding. On a busy day it can use around 5kWh, though on most days it is around 2kWh. I am sure cloud is more efficient because there is probably efficiency in running many parallel streams, some of the models have fever than 27B active parameters and the power draw of the non-GPU components is also spread over more GPUs. But still I also believe that the energy use numbers you get from the inference providers are a bit "beautified". After all, they still fight a political battle and have to show that it isn't all so bad.
Having said that, seeing the incredible progress of models throughout this year, I also strongly believe that the planned buildout is overeager. Even I with my gaming hardware often run out of instructions to give. And the smarter the models get that I can run, the less I will be able to saturate my hardware. Is it because of my lack of creativity of which kinds of tasks I can give to AI? Maybe a bit, but currently I can't believe that I am that far away.","created_at":"2026-10-03T07:04:31Z","created_at_i":1791011071,"objectID":"49942000","parent_id":49937645,"story_id":49934620,"story_title":"One month coding with GLM 5.3 Flash","story_url":"https://wagtail.org/blog/one-month-on-glm-53-flash/","updated_at":"2026-10-13T08:37:54Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"KiwiJohnno"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["qwen","coding","agent"],"value":"Yeah before Qwen 3.8 27b all local LLMs did feel like not very smart chat completion models. You would hit the limits of their intelligence all the time. Chatting with Qwen 3.8 is night and day, and it can do some pretty damn good coding and agentic work for its size. IMO its better than where frontier labs were at about 18 months ago, which is mind-blowing for a small locally hosted model."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Several vulnerabilities have been discovered in the Linux kernel"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://lwn.net/Articles/1097401/"}},"_tags":["comment","author_KiwiJohnno","story_49928121"],"author":"KiwiJohnno","comment_text":"Yeah before Qwen 3.8 27b all local LLMs did feel like not very smart chat completion models. You would hit the limits of their intelligence all the time. Chatting with Qwen 3.8 is night and day, and it can do some pretty damn good coding and agentic work for its size. IMO its better than where frontier labs were at about 18 months ago, which is mind-blowing for a small locally hosted model.","created_at":"2026-10-03T01:10:35Z","created_at_i":1790989835,"objectID":"49940480","parent_id":49934403,"story_id":49928121,"story_title":"Several vulnerabilities have been discovered in the Linux kernel","story_url":"https://lwn.net/Articles/1097401/","updated_at":"2026-10-03T01:16:20Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"desmondl"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["qwen","coding","agent"],"value":"Yeah, I clicked expecting a review of GLM 5.3 Flash, but the article was more a retrospective about what he learned during his September challenge: "Only use GLM 5.3 Flash for one month"
He said his experiment was a failure because:
1. He accidentally spent 450M tokens vibe coding with the wrong model, instead of GLM 5.3 Flash.
2. When he used GLM 5.3 Flash, it was sometimes slow. So he switched to other models (Deepseek / Qwen) instead. His guess to why it was slow: GLM 5.3 Flash was so good that the providers were congested.
3. He still needed to use other models besides GLM 5.3 Flash, for R&D and benchmarking.
His takeaways from doing the experiment were:
1. Measure local usage more.
2. Experiment with agent orchestration, with bounded goals.
3. Don't count other models that are used for R&D.
4. Play with Jev.
5. Include experiments with flagship models to compare with cheap open models.
His conclusion about GLM 5.3 Flash: Probably viable for day to day work, but he'll have more thoughts next month."},"story_title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["coding"],"value":"One month coding with GLM 5.3 Flash"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://wagtail.org/blog/one-month-on-glm-53-flash/"}},"_tags":["comment","author_desmondl","story_49934620"],"author":"desmondl","children":[49942661],"comment_text":"Yeah, I clicked expecting a review of GLM 5.3 Flash, but the article was more a retrospective about what he learned during his September challenge: "Only use GLM 5.3 Flash for one month"
He said his experiment was a failure because:
1. He accidentally spent 450M tokens vibe coding with the wrong model, instead of GLM 5.3 Flash.
2. When he used GLM 5.3 Flash, it was sometimes slow. So he switched to other models (Deepseek / Qwen) instead. His guess to why it was slow: GLM 5.3 Flash was so good that the providers were congested.
3. He still needed to use other models besides GLM 5.3 Flash, for R&D and benchmarking.
His takeaways from doing the experiment were:
1. Measure local usage more.
2. Experiment with agent orchestration, with bounded goals.
3. Don't count other models that are used for R&D.
4. Play with Jev.
5. Include experiments with flagship models to compare with cheap open models.
His conclusion about GLM 5.3 Flash: Probably viable for day to day work, but he'll have more thoughts next month.","created_at":"2026-10-02T21:18:21Z","created_at_i":1790975901,"objectID":"49938713","parent_id":49937824,"story_id":49934620,"story_title":"One month coding with GLM 5.3 Flash","story_url":"https://wagtail.org/blog/one-month-on-glm-53-flash/","updated_at":"2026-10-07T07:39:34Z"}],"hitsPerPage":20,"nbHits":506,"nbPages":26,"page":0,"params":"query=qwen+coding+agent&advancedSyntax=true&analyticsTags=backend","processingTimeMS":11,"processingTimingsMS":{"_request":{"roundTrip":20},"afterFetch":{"format":{"highlighting":1,"total":1}},"fetch":{"query":6,"scanning":3,"total":10},"total":11},"query":"qwen coding agent","serverTimeMS":13}