End of September 2026 recommendations are here: https://aibynick.com/thread/80

Many valid details are also still available in the August-mid September 2026 roundup: https://aibynick.com/thread/63

For the overwhelming majority of daily work, I currently use Deepseek V4.1 Flash on the Cline-pass API (https://cline.bot/cline-pass). That account, which costs less than $7 per month, is one of the most valuable services in my life. I build most of the software I deliver to clients, complete IT work (installations, OS configs, compilations, etc.), and complete my own personal tasks with it, all day every day.

For any workflow that involves PHI or compliance obligations, that's where my self-hosted models are required (personal health info can't be legally sent to most of the public LLM API providers). I run GLM 5.3 flash iq3_xxs on a cluster of 2 DGX Sparks, and Mimo 2.6 Flash MXFP4 on a separate cluster of 2 DGX sparks. Those jobs most often involve working with private account credentials, building software which can't be shared publicly, and anonymizing data in various file formats (so I can work with the shape of that de-identified fake data in faster APIs such as Cline-pass).

Qwen 3.8 Flash Next iq4_xs currently tends to stay loaded on my Strix Halo machines. It's amazingly capable for its size, it's fast, full of world knowledge, able to understand and quickly reason through significantly challenging problems. To be honest, I could likely do everything I need with only that model, if I had to. GLM and Mimi are even more capable, so they require less hand holding and baby-step iterations to complete development cycles, but with a little elbow grease, QFN is capable of replacing lots of infrastructure and expenses.

Qwen 3.6 35a3 q4 runs most often on my 3 machines which have RTX GPUs with 16-24GB VRAM. That old model is fantastically capable, for a tiny, incredibly quick model. Qwen 3.8 27b is the reigning popular champ, but I've had better success with 3.6 35a3 for the sorts of tasks I hand it. That's an unpopular opinion, but it's much faster and has been more reliable for my purposes (likely because I manually take part in more engineering steps than non-technical users prefer to, or are able to).

I keep a Deepseek API key handy for quick configs and for use in temporary environments where Cline-pass involves more hurdles to set up. I also use OpenRouter, especially to test new models. Using those providers for extended daily work tasks, however, is simply a waste of money, compared to using Cline-pass.

For the overwhelming majority of my daily work, I run every model in Pi coding agent, on local machines (even on my Android phone, in Termux), and on several VPS accounts hosted at Contabo ($5 per month). Locally, I run it on the command line. On the VPSs, I use the Pi-web interface, which lets me connect and run it in any browser, on any device with Internet access. Because of Pi-web, I've begun to use a single Pi installation in one VPS account, most often. That gives me access to the same active session workflows, files, installed software, etc., from any location, using an machine.

I can use any of the models, from any of the LLM API provider accounts, or any of my self-hosted models, in any of the Pi installations (and those models can all be immediately hot swapped if needed) - though Deepseek v4.1 Flash is really the only default model I use constantly these days. It's a phenomenally capable multimodal LLM, which typically runs 200-250 tokens per second on the Cline-pass API (I've seen it exceed 300 tokens per second). That's outrageously fast for such an inexpensive API service.

It's finally getting to the point that the combination of the DS4.1f model, Pi harness, Pi-web UI, Cline-pass LLM API (+ self-hosted models), and VPS hosting are replacing all my other workflow tooling, including the ChatGPT zip file workflow which has been the cornerstone of my software development work for more than a year. I still keep updating legacy projects in ChatGPT, but the new pi-web environment works remarkably well for new projects, and it's no biggie to switch any ChatGPT project over to pi-web, whenever needed.

Finally, to manage a fleet of old 32-bit Windows 7 machines, I recently compiled a version of Nullclaw (https://aibynick.com/thread/88). That harness makes those ancient machines capable of completing all sorts of useful tasks, with the same effortless agentic workflow we've gotten used to using on new machines. Since RAM and other components are currently so expensive, those old machines are a potential solution for people and organizations who don't have piles of cash to burn on new hardware. An $80 per year Cline-pass account, and a used $100 netbook are all you need to complete lots of useful work. Add a $50 per year VPS account, and you can deliver production solutions.

End of September 2026 recommendations are here: https://aibynick.com/thread/80

Many valid details are also still available in the August-mid September 2026 roundup: https://aibynick.com/thread/63

The way it currently stands, for the overwhelming majority of daily work, I use Deepseek V4.1 Flash on the Cline-pass API. That account, which costs less than $7 per month, is one of the most valuable services in my life. I build most software and complete IT work (installations, OS configs, compilations, etc.) with it all day every day.

For any task that involves PHI or compliance obligations, that's where my self-hosted models are required. I run Mimo 2.6 Flash MXFP4 on a cluster of 2 DGX Sparks, and GLM 5.3 flash iq3_xxs on a separate cluster of 2 DGX sparks. Those jobs most often involve working with private account credentials, building software which can't be shared publicly, and anonymizing data in various file formats (so I can work with the shape of that data in faster APIs such as Cline-pass).

Qwen 3.8 Flash Next iq4_xs currently tends to stay loaded on my Strix Halo machines. It's amazingly capable, fast, full of world knowledge, able to understand and reason through challenging problems quickly, etc. To be honest, I could likely do everything I need with only that model, if I had to. GLM and Mimi are even more capable, so they require less hand holding and iterations, but with a little elbow grease, QFN is capable of replacing lots of infrastructure and expenses.

Qwen 3.6 35a3 q4 runs most often on my 3 machines which have RTX GPUs with 16-24GB VRAM (that old model is fantastically capable, for a tiny, incredibly fast model).

I also keep a Deepseek API key handy for quick configs and for use in temporary environments where Cline-pass takes longer to set up, and I use OpenRouter to test new models.

For the overwhelming majority of my daily work, I run every model in Pi coding agent, on local machines (even on my Android phone, in Termux), and on several VPS account hosted at Contabo ($5 per month). Locally, I run it on the command line. On the VPSs, I use the Pi-web interface, which lets me connect and run it in any browser, on any device with Internet access. Because of Pi-web, I've begun to use the Pi installation in the VPS account most often.

I can use any of the models, from any of the LLM API provider accounts, or any of my self hosted models, in any of the Pi installations - though Deepseek v4.1 Flash is really all I default to constantly these days.

It's finally getting to the point that that combination of ds41f model, pi harness, pi-web UI, LLM API, and VPS hosting are replacing all my other workflows, including the ChatGPT zip file workflow which has been a cornerstone of my development work for more than a year (I still keep all the legacy projects in ChatGPT, but the new pi-web environment works better for new projects, and it's no biggie to switch and ChatGPT project over to pi-web, if needed).

To manage a fleet of old 32-bit Windows 7 machines, I compiled a version of Nullclaw (https://aibynick.com/thread/88). That harness makes those ancient machines capable of completing all sorts of useful tasks, with the same effortless agentic workflow we've gotten used to using on new machines. Since RAM and other components are currently so expensive, those old machines can be a solution for people and organizations who don't have piles of money to spend on hardware. $80 per year Cline-pass account is all you need.

End of September 2026 recommendations are here: https://aibynick.com/thread/80

Many valid details are also still in the August-mid September 2026 roundup: https://aibynick.com/thread/63

The way it currently stands, for the overwhelming majority of daily work, I use Deepseek V4.1 Flash on the Cline-pass API. That $7 per month account is one of the most valuable services in my life. I build most software and complete IT work (installations, OS configs, compilations, etc.) with it all day every day.

For any task that involves PHI or compliance obligations, that's where my self-hosted models are required. I run Mimo 2.6 Flash MXFP4 on a cluster of 2 DGX Sparks, and GLM 5.3 flash iq3_xxs on a separate cluster of 2 DGX sparks. Those jobs most often involve working with private account credentials, building software which can't be shared publicly, and anonymizing data in various file formats (so I can work with the shape of that data in faster APIs such as Cline-pass).

Qwen 3.8 Flash Next iq4_xs currently tends to stay loaded on my Strix Halo machines. It's amazingly capable, fast, full of world knowledge, able to understand and reason through challenging problems quickly, etc. To be honest, I could likely do everything I need with only that model, if I had to. GLM and Mimi are even more capable, so they require less hand holding and iterations, but with a little elbow grease, QFN is capable of replacing lots of infrastructure and expenses.

Qwen 3.6 35a3 q4 runs most often on my 3 machines which have RTX GPUs with 16-24GB VRAM (that old model is fantastically capable, for a tiny, incredibly fast model).

I also keep a Deepseek API key handy for quick configs and for use in temporary environments where Cline-pass takes longer to set up, and I use OpenRouter to test new models.

For the overwhelming majority of my daily work, I run every model in Pi coding agent, on local machines (even on my Android phone, in Termux), and on several VPS account hosted at Contabo ($5 per month). Locally, I run it on the command line. On the VPSs, I use the Pi-web interface, which lets me connect and run it in any browser, on any device with Internet access. Because of Pi-web, I've begun to use the Pi installation in the VPS account most often.

I can use any of the models, from any of the LLM API provider accounts, or any of my self hosted models, in any of the Pi installations - though Deepseek v4.1 Flash is really all I default to constantly these days.

It's finally getting to the point that that combination of ds41f model, pi harness, pi-web UI, LLM API, and VPS hosting are replacing all my other workflows, including the ChatGPT zip file workflow which has been a cornerstone of my development work for more than a year (I still keep all the legacy projects in ChatGPT, but the new pi-web environment works better for new projects, and it's no biggie to switch and ChatGPT project over to pi-web, if needed).

To manage a fleet of old 32-bit Windows 7 machines, I compiled a version of Nullclaw (https://aibynick.com/thread/88). That harness makes those ancient machines capable of completing all sorts of useful tasks, with the same effortless agentic workflow we've gotten used to using on new machines. Since RAM and other components are currently so expensive, those old machines can be a solution for people and organizations who don't have piles of money to spend on hardware. $80 per year Cline-pass account is all you need.

The way it currently stands, for the overwhelming majority of daily work, I use Deepseek V4.1 Flash on the Cline-pass API. That $7 per month account is one of the most valuable services in my life. I build most software and complete IT work (installations, OS configs, compilations, etc.) with it all day every day.

For any task that involves PHI or compliance obligations, that's where my self-hosted models are required. I run Mimo 2.6 Flash MXFP4 on a cluster of 2 DGX Sparks, and GLM 5.3 flash iq3_xxs on a separate cluster of 2 DGX sparks. Those jobs most often involve working with private account credentials, building software which can't be shared publicly, and anonymizing data in various file formats (so I can work with the shape of that data in faster APIs such as Cline-pass).

Qwen 3.8 Flash Next iq4_xs currently tends to stay loaded on my Strix Halo machines. It's amazingly capable, fast, full of world knowledge, able to understand and reason through challenging problems quickly, etc. To be honest, I could likely do everything I need with only that model, if I had to. GLM and Mimi are even more capable, so they require less hand holding and iterations, but with a little elbow grease, QFN is capable of replacing lots of infrastructure and expenses.

Qwen 3.6 35a3 q4 runs most often on my 3 machines which have RTX GPUs with 16-24GB VRAM (that old model is fantastically capable, for a tiny, incredibly fast model).

I also keep a Deepseek API key handy for quick configs and for use in temporary environments where Cline-pass takes longer to set up, and I use OpenRouter to test new models.

I run every model in Pi coding agent, on local machines (even on my Android phone, in Termux), and on several VPS account hosted at Contabo ($5 per month). Locally, I run it on the command line. On the VPSs, I use the Pi-web interface, which lets me connect and run it in any browser, on any device with Internet access. Because of Pi-web I've begun to use the Pi installation in the VPS account most often.

I can use any of the models on any of the API accounts, or any of my self hosted models, in any of the Pi installations, though Deepseek v4.1 Flash is really all I default to constantly these days. It's finally getting to the point that that model, harness, web UI, LLM API, and VPS hosting are replacing all my other workflows, including the ChatGPT zip file workflow (I still keep all the legacy projects in ChatGPT, but the new pi-web environment works better for most new projects, and it's no biggie to switch and ChatGPT project over to pi-web).

To manage a fleet of old 32-bit Windows 7 machines, I compiled a version of Nullclaw: https://aibynick.com/thread/88 . That makes those ancient machines capable of completing all sorts of useful tasks, with the same effortless agentic workflow we've gotten used to on new machines. Since RAM and other components is currently so expensive, those old machines can be a solution for people and organizations which don't have piles of money to spend on hardware.