Post History

Current VersionJul 06, 2026 at 00:58

I've been using Deepseek V4 Pro on Openrouter as my main LLM workhorse for agentic work, ever since it came out. Pi has been my workhorse harness.

Recently, I've begun using v4 Flash more than v4 Pro as my daily LLM driver. The Flash version provides an especially practical mix of capability, speed, and low cost - I'm beginning to trust it for more and more all sorts of work.

My initial proving ground for Deepseek V4 Pro and Pi was a tough project I'd been working on for several months, building a bot to automate interaction with 3rd party dashboards, using natural language queries from users, accessed via a web UI. The project involved deeply crawling the dashboards, using Playwright code to perform UI interactions with their JS-heavy interfaces.

The project required a full crawl of every single UI widget on every page of each dashboard application, complete with generated documentation of every data selection available in every widget (for example, all the values in every dropdown selector, and every text field that could be entered by the user, along with the sorts of values that a user is expected to enter in those fields). The crawling routines spidered each of the dashboard sites to create a full knowledge base which contained all the logic and recipes that a public facing agent would need, in order to understand where to look within the dashboards, for answers to any sort of natural language questions a user could potentially query at the public interface.

The crawling portion of the project required a massive volume of Playwright code to be written by an agent, to discover and document every single UI control which existed in several dashboards. That was a tremendous volume of work.

I started that project in January, using Claude Code and the various Claude LLMs. That first set of tools was successful at writing some functional Playwright code which could interact with a limited set of specific controls in 1 dashboard, but the rate limits imposed by Anthropic were debilitating. We could only run a handful of interactions each day before the API hit its limit, and the process was slow.

Next in that project, I migrated everything we had accomplished with Claude Code, to use the Goose agent instead. That change enabled us to use any LLM available on Openrouter. The project really moved forward with the help of a variety of Gemini models that didn't have rate limits. Gemini 3.1 Flash Light Preview was particularly fast, but not quite smart enough to get the job done. It was incredible to watch that model burn tens of millions of tokens overnight, unattended, in Nullclaw, for just a few dollars. But still, getting a fully working app together was a mess, because the fastest/cheapest sub-frontier class flash models still weren't quite smart enough to understand and engineer everything required to achieve our goal, and the smarter models were far to expensive to run throughout the huge crawling routine. Using the best Claude and GPT models to perform the massive volume of crawling work would have cost 10s of thousands of dollars - beyond the budget of a relatively small research grant which was driving the project.

Once Deepseek v4 Pro and Pi became my goto model and harness, this project moved forward incredibly quickly and easily. Crawling was easily completed in just a few days, for a total cost of about $250 for all the LLM work.

Deepseek V4 Pro not only automatically wrote all the Playwright code needed to perform all the crawling tasks - live in unattended sessions - it also wrote all the public facing web application code (along with writing and performing the full suite of tests required to build the app). It autonomously performed the full crawling portion of the project, which required a lot of reasoned decisions, intuitive understanding of the sorts of natural language questions users might ask, etc. - all of which together, was a staggering long horizon feat that it completed with ease.

During that period, I also began using Deepseek v4 Pro to complete all my other production development efforts which weren't being handled in ChatGPT zip file workflows (see https://aibynick.com/thread/3). I additionally produced heaping piles of demo applications & games, and performed every imaginable computer control agentic task using Deepseek v4 Pro. It became my unrivaled go-to for local agentic work, without breaking the bank.

Around that time, Qwen 3.6 was released, and it became my default go-to for inference on local LLM servers. Those Qwen models are brilliantly capable at writing code, for the minimal hardware they require - but they don't have the world knowledge and deep intellectual capability that frontier models have.

So, I'm very comfortable giving CRUD database and basic UI development work, as well as simple agentic tasks, to every common quant of Qwen 3.6 35a3 and 27b. And Gemma 4 has been useful for some tasks which require vision. But when I need to do real engineering for bigger projects, I considered Deepseek v4 Pro the minimum model to employ.

Keep in mind that most of my big software development projects are still typically accomplished with the ChatGPT zip file workflow, since I get all my big projects done with it for $20/month, total - and GPT 5.5 is just so smart at understanding goals, and reasoning its way through complex engineering plans. I can often just give ChatGPT the requirements that my clients send by email, and software updates get built and tested automatically, in a few minutes, first shot, with GPT 5.5. As long as they're giving away all that intellectual power for $20, I'll be using it relentlessly.

Still, I know that unlimited subsidized GPT tokens for $20/month will be going away as soon as those OpenAI investor dollars get burned up, and I also do find myself needing local AI to deal with tasks that involve HIPAA compliance (I can't put PHI into any typical public LLM API). Additionally, my clients are starting to ask for self-hosted HIPAA compliant LLM API's, because services such as Bastion GPT cost a lot of money per million tokens.

So, local AI inference is becoming a much more critical priority for me and my clients. I've already gotten used to using a local collection of 3 Asus GX10s, 2 Strix Halo machines, and 3 other lesser machines with 3080 and 3090 class GPUs, every day to compete tasks.

The problem has been that not every task can be completed by small models like Qwen 3.6, and larger models like Deepseek v4 Pro require 10s of thousands of dollars of hardware to run. Most of my clients are quite ready to take that plunge.

Minimax M3 is a real contender for serious self-hosted LLM inference, but it still requires at very least 3 clustered DGX Sparks to run reliably - and Minimax Sparse Attention (which is required for usable million context window size) still isn't available in the GGUFs from Unsloth or other quants.

That's why my interest has turned solidly to Deepseek V4 Flash. It's possible to run a low bit quant of V4 Flash on a single DGX Spark, and a truly reliable quant can be run on a cluster as small as 2 Strix Halo machines. See this video:

https://www.youtube.com/watch?v=Cfl3TS7ME5s

The quality of agentic task results demonstrated by Deepseek v4 Flash in that video are amazingly good: 90% on the Dwarfstar benchmarks (start watching at time stamp 16:48 in the video above to see those results).

On 2 DGX Spark machines, Deepseek V4 Flash is a natural fit, with expected speeds that are much faster than on clustered Strix Halo machines, due to the much faster native NVIDIA ConnectX-7 network connectors built into that platform.

So, I've been using V4 Flash to do a ton of daily driving lately, and it is remarkably capable, especially at typical daily coding, agentic tasks, reasoning, etc. And on Openrouter and elsewhere, the cost of V4 Flash is less that 1/4 that of V4 Pro - and it's ridiculously fast.

I've been using V4 Flash everywhere for the past week, for some significantly challenging work, and it has not needed the help of any bigger frontier model, at any point.

That really changes the game, for many reasons.

Not only is V4 Flash ridiculously cheap and capable, it can also be run locally. That makes all the difference for my clients who require complete control over data security.

I started my exploration of V4 Flash by building the same sorts of little vibe coded demos I do with every model - some quick CRUD apps such as a standard Northwind database flask example, some small games, 3D examples, website UI layouts, etc. For example, here are the results of just a few quick coding tests:

Here are a few more examples made by the Deepseek-v4-flash model - its really amazing how inexpensive and capable that model is!:

What shocked me was that many of the examples cost less than a penny to create (including a few of the games above), they were all completed in record time, and every example worked perfectly, first shot, out of the gate.

That was immediately impressive, but not necessarily representative of how well the model would handle challenging real life tasks. So I've been increasing the complexity dramatically, and V4 Flash has not let me down.

For example, today, I completed a personal project with V4 Flash, which I've been wanting to do for a while now. The goal was to set up voice control of Pi in Termux (Linux) on my Android phone, to interact with Termux API to provide device control of all typical Android functionalities.

The results of that work today, were mind blowing. Deepseek V4 Flash was able to write an extension module for Pi, to process voice prompts using the Android voice API (it's not otherwise possible in Pi, within Termux, to use native Android voice dictation).

Then Deepseek V4 Flash wrote all the code needed to interact with Termux API, which gives Termux the ability to control other native Android functionalities such as making calls, sending texts, accessing GPS and hardware sensors, etc. And of course, I completed all that work using my new Pi voice interface, directly on my phone :) (Beware that you need Termux and Termux API downloaded directly from F-droid, because Termux API does not work with the version from Google Play Store).

The combination of those two new features was really exciting. Now I can open Pi on my phone and use voice dictation to do virtually anything the device is capable of. I had V4 Flash update Pi's .md files/memory to codify everything that was learned about how to interact with voice inputs and the Termux API.

I subsequently had a blast telling Pi, on my phone, entirely with natural language, to perform tasks such as 'take a photo of me and share it with my girlfriend after I view and approve it' (my Pi install can perform the photo task because Termux API provides control of things like the camera and photo apps on Android). Then I told Pi to perform online research about one of my client's company, and to send them a few texts about the results, along with some texts I had it compose to explain what I was doing.

This sort of unbounded 'computer control' on my phone, matched with full natural voice interaction is truly useful stuff - it really opens up a whole new world of usability and functionality on my phone which enables me to be productive in ways that I've dreamed about for decades. The net effect is that I can just talk to my phone, to have it accomplish basically any sort of work. I could realistically have it complete large projects, without ever having to type a character. That's a real step change in the functional capability of my phone.

Deepseek V4 Flash completely crushed doing all the research, writing all the code, interacting with my development requests, and then running my queries. All the code was written flawlessly, at blindingly fast speed, first shot.

Now here's the kicker: all that work today required burning ~8.5 million tokens on Openrouter - and the entire cost for everything - all of the development work and all the interactivity involved in using the features I had built... cost $.34.

34 cents

That is genuinely game-changing.

For my clients, the v4 Flash model opens up the opportunity to buy some very reasonably priced local hardware - 2 Asus GX10s, for example - and to have powerful enough LLMs available to work with PHI in agentic tasks, entirely in-house. Those clients would otherwise spend far more on HIPAA compliant LLM API services such BastionGPT, in less than a year.

And of course, then there are so many other practical applications. Deepseek 4 is not multimodal, but that's fine. I've already seen models such as Qwen 3.6, Gemma 4, and Stepfun Flash 3.7 do a great job identifying wound images, for example.

Mimo-v2.5 is looking like another very capable, similarly inexpensive vision model which can also run on systems like those same clustered DGX Sparks. Mimo-v2.5 is likely my next medium-ish sized model to really put through the paces. It's impressive for a size that can be legitimately hosted locally.

Going forward, Deepseek V4 Flash will be my default choice for on-device agentic tasks, and likely for most small-to-medium sized software development tasks. It's so outrageously inexpensive for the capability it provides, it's fast, and it can be self-hosted. And if it ever needs help, I can just switch to another model at any point in any development process.

I love the idea of really getting to know exactly what I can expect of V4 Flash, in any situation - and then knowing which models are needed to replace it, for anything that it can't handle immediately out of the gate.

I can truly see a light at the end of tunnel, in which frontier models may not be required at all, for most typical commercial development work and agentic workflows.

Version 8Jul 06, 2026 at 00:58