IT (dis)adventures: DeepSeek

For two (three?) years, serious agentic coding meant paying frontier prices (and keep increasing!!) to a few closed labs. DeepSeek 4, and V4.1 Flash in particular, suggests that rule is about to break.

For the past two years, agentic software development has followed a simple rule: if you wanted an AI that could reliably navigate a codebase, run commands, fix failing tests and iterate without supervision, you paid frontier prices to a handful of American labs. Cheaper and open-weight models existed, but they fell apart on exactly the long, messy, multi-step tasks that make coding agents useful.

That rule may be breaking. The next few months could be remembered as the moment agentic coding stopped being a premium service and became a commodity.

A Flash model that beat its own flagship

The story starts with DeepSeek V4, and it accelerated quickly. At the end of July, DeepSeek promoted V4-Flash from preview to official release, and a model with 13B activated parameters beat the 49B-active V4-Pro preview on all nine agentic benchmarks DeepSeek publishes. On DeepSeek's own table, that 0731 build was already close to Anthropic's previous flagship: 82.7 on Terminal Bench 2.1 against 85.0 for Opus 4.8, and 54.4 on DeepSWE against 58.0.

I'm currently

Then came V4.1 Flash. It shipped on 10 September 2026, and DeepSeek is retiring its own V4 Pro because the smaller model beats it on most tests. According to DeepSeek, at maximum reasoning effort V4.1 Flash outperformed OpenAI's GPT-5.6 Sol and Anthropic's Claude Opus 5 on Terminal-Bench 2.1. The headline numbers: Terminal-Bench 2.1 at 90.6, DeepSWE at 74.2, a Codeforces rating of 3471, and HLE-with-tools at 63.9, just past Claude Opus 5's 63.6.

I've been working on it for a week now, on a .NET 10 Core solution with more than 10,000 C# files, trying to push it further and further, and I have to say I'm honestly amazed, especially considering the cost! I don't want to advertise, but I'm currently using it through Ollama cloud, which costs less than the Deepseek API. I've also read that OpenCode cloud offers it at an even lower cost, but I haven't tried it yet because I prefer using Visual Studio 2026 or VS Code, at least for now.

What is interesting is where the gains came from. The model card describes a standard post-training recipe (SFT, RL and on-policy distillation) with no algorithmic changes; the substantive work went into the data pipeline, with large-scale automated synthesis of agent tasks and environments. In other words, agentic competence is turning out to be something you can train into a relatively small model, not something that requires trillion-parameter scale. That is a finding with consequences for the entire industry.

The price gap is not a gap, it is a canyon

Benchmarks alone would make this a good release. The pricing makes it potentially disruptive. Off-peak, V4.1 Flash costs $0.15 per million input tokens and $0.60 per million output, with peak hours doubling those rates. For comparison, Claude Opus 5 costs $5 per million input tokens and $25 per million output.

That is roughly a 20x to 40x difference per token, depending on the time of day. And agentic coding is precisely the workload where per-token cost matters most, because every tool result goes back into the context and the growing transcript is billed again on each turn. The architecture helps here too: a 20-layer encoder reads the input with just 8 billion active parameters, a 20-layer decoder writes the reply with 16 billion, and the KV cache shrinks to 890 bytes per token, about a quarter of V4 Flash's. Agent work is input-heavy, and this model is built to read cheaply.

When a task that cost ten dollars costs thirty cents, the economics of how you use agents change qualitatively. You stop rationing. You run five attempts in parallel and keep the best one. You let an agent grind on a refactor overnight. You put an agent on every pull request instead of only the ones that seem important.

The part nobody could do before: owning the model

Price is one axis. Control is the other, and this is where things get genuinely new. DeepSeek releases its weights under an MIT license, which means a near-frontier coding model is something you can run on hardware you own, with no API, no data leaving your network, and no provider swapping the model under you.

This is where Salvatore Sanfilippo, the creator of Redis, enters the picture. His project DwarfStar (ds4) aims to be the best way to run a few excellent large language models on consumer hardware, meaning hardware people can actually own, through a small native inference engine optimized first for DeepSeek V4 Flash and V4.1 Flash. It is deliberately narrow: rather than being a general-purpose runner, it specializes in a handful of models and tests everything together, from tool calls to KV state to a built-in coding agent. It targets Metal on Macs with 96 GB or more, NVIDIA CUDA with the DGX Spark as its main goal, and ROCm on Strix Halo systems such as the Framework Desktop.

The practical results are striking. For V4 Flash, a first run on a 96 or 128 GB machine starts with the Q2 quantization, and SSD streaming lets even machines without enough RAM run it at a decent speed. The heavier V4.1 Flash is more demanding: Q2 runs with SSD streaming on a single 128 GB Mac or Spark, or fully resident across two Macs or two Sparks linked via RDMA. The included server also works with Pi, OpenCode, Codex CLI or Claude Code, so developers can keep their existing tooling and swap the brain underneath.

There is a meta-lesson in the project itself. Sanfilippo is open that ds4 is developed with strong assistance from AI coding agents, with humans leading the ideas, testing and debugging. He even suggests users treat coding agents as the interface to the project, asking them to tune performance for their specific hardware. The tool for running local coding agents is itself a product of coding agents. That feedback loop is exactly what becomes explosive once the agents are cheap and local.

Why the next few months matter

Put the pieces together. A near-frontier agentic coding model exists. It costs a small fraction of the frontier per token. Its weights are free to download. And there is now a focused, well-engineered path to running it on a single high-end workstation that a freelancer or a small company can buy outright.

Any one of those facts would be notable. Together they threaten the business model that has defined the field: the assumption that serious agentic coding means a recurring subscription to a closed lab. Expect startups to rebuild their agent products on open weights, enterprises with strict data policies to finally approve coding agents because the model never leaves the building, and frontier labs to face real pressure on pricing for the "workhorse" tier of their lineups.

The honest caveats

It would be a mistake to declare the frontier dead. Most of these numbers are vendor-reported, and the picture changes on harder tests: V4.1 Flash drops to 30.0 on Terminal-Bench 3.0, handling standard terminal tasks well but struggling with the more demanding agentic suite. On pure reasoning without tools it also trails: on Humanity's Last Exam it scores 36.8, well under GPT-5.6 Sol at 44.5 and Claude Opus 5 at 56.3. Even its better results leave room for doubt; one analysis notes that a 54.8 on Automation-Bench means it still fails roughly half of complex workflows.

There are practical issues too. The hosted API raises data-governance questions for some organizations, and local hardware capable of running the larger model is still a serious investment. And the frontier labs will not stand still.

But revolutions rarely require the challenger to be better. They require it to be good enough, much cheaper, and in the hands of more people. For agentic software development, DeepSeek 4 and projects like DwarfStar suggest we may be crossing that threshold right now.

A European view: the geopolitics of open weights

There is an uncomfortable question hanging over all of this, and for a European reader it cannot be avoided: should we be building on Chinese models at all?

Let us start from reality rather than wishful thinking. Europe sits firmly within the American sphere of influence, and for the foreseeable future that is not going to change. Our security architecture runs through NATO, our cloud runs largely on American hyperscalers, our phones, operating systems and productivity suites are American, and the frontier AI labs we rely on are American. Talk of "digital sovereignty" is valuable as an aspiration, but Europe does not currently have the capital, the compute or the industrial scale to replace that ecosystem. Being part of the Western bloc is, at least for now, a given.

Accepting that, however, does not mean Europe should treat every technology coming from China as radioactive. The key is to distinguish between two very different things that often get lumped together under the same name.

A Chinese service is not the same as a Chinese model

Using a Chinese service means sending your prompts, your source code and your data to servers in the People's Republic, subject to Chinese law and outside any European oversight. That is a legitimate concern, and it is the reason several European regulators, starting with Italy's data protection authority, moved against DeepSeek's consumer app early on.

Running Chinese open weights is a different proposition entirely. The weights are a file. Once downloaded and executed on a workstation in Perugia, a server in Frankfurt or a GPU cluster run by a European or American provider, the model has no way to phone home, no account tied to Beijing, and no data flow leaving the infrastructure you control. The MIT license means no one can revoke your right to use it or change it under your feet. In a very literal sense, an open-weight model running locally is more under your control than any closed Western model you access through an API.

This is precisely why tools like DwarfStar matter beyond the technical achievement. They turn a model trained in China into a piece of European infrastructure, running on hardware owned by Europeans, governed by European law.

Pragmatism, not alignment

Refusing to use open Chinese models out of principle would not make Europe more Western; it would only make it poorer and more dependent. It would mean paying frontier API prices for work that a near-equivalent model could do on our own machines, and it would mean giving up one of the few levers that actually strengthens European autonomy: the ability to run capable AI without asking anyone's permission.

There is also a competitive argument. Open weights from DeepSeek, Qwen and others put pressure on prices across the whole market, including on American providers. Europe benefits from that competition as a customer, and it benefits even more if its own companies and researchers learn from, fine-tune and build on top of these models instead of watching from the sidelines.

The objections worth taking seriously

None of this means the risks are imaginary. Weights are opaque: a model can carry biases or subtle behaviours from its training that are hard to detect, and Chinese models are known to follow the official line on politically sensitive topics. Some security researchers also worry about models being trained to insert weaknesses into generated code, a risk that is hard to rule out for any model, but one that a geopolitical rival has more motive to exploit. And there is a political dimension: in a period of tension between Washington and Beijing, European companies that depend heavily on Chinese technology could find themselves caught in future trade or export restrictions.

These concerns argue for caution, not for a ban. The sensible response is the same one we should apply to any model, Western or not: run it on infrastructure we control, keep humans reviewing what goes into production, test it independently, and never let a single supplier become irreplaceable. Used this way, open Chinese models are not a threat to Europe's place in the West. They are one more tool that lets Europe be a little less dependent on everyone.

Sources

IT (dis)adventures

Friday, September 18, 2026

The Cheap Model That Might Change Everything: What Happens When "Good Enough" Becomes "Almost Best"