DeepSeek V4.1-Flash-1017 is INSANE!

Julian Goldie (0:00)

So, DeepSeq V4.1-Flash-1017 is here. Stronger benchmark performance, faster task completion, shorter outputs. This just dropped, this literally just dropped a few hours ago, as you can see. One of the important things to note here is this is a Flash model, so it's very fast. But at the same time, it is competing with frontier-level technology, really powerful new API that's just dropped. It can build some absolutely awesome stuff. And also, it's available already on the API. And additionally, you can see that it's open weights. Let's talk through what this means and why it's different for everyone watching today. All right, so this just dropped today, DeepSeek V4.1-Flash-107.

And there's 11 benchmark, 6 models, and one very cheap price tag that's involved in this story. So let's get straight into this. So first of all, this is the same name, and it's the same price, but it's a noticeably sharper model than previous versions of DeepSeek. So 1017 isn't a new generation. It's just a fresh post-training checkpoint of September's V4.1-Flash, and the date is the name, right? 1017, so 1017 So DeepSeek have actually swapped it in on the API, the website, the apps, and this is basically the new update on Flash. Now, what DeepSeek calls this is Pareto Improvement, which means better scores, is faster, shorter answers with nothing traded away, and the training emphasis was reinforcement learning and whole end-to-end projects from building games to 3D modeling. So it's much better coding, it's much better at visual stuff. And one thing to note here is 522 billion parameters plus Ngram memory parameters, 196 billion. And it's a mixture of experts model, so 8 billion parameters active are actually switched on per token, 1 million token context window, maximum 384k output tokens in a single response, 890 bytes. So that's the memory per token in the main attention cache. And 1 to 100 reason and effort dial on the API. You can see the cost of the API right here, and it's MIT license. So this is an open weight model already available on Hugging Face, like you can see right here. Happy days. And you might be wondering, okay, what's the point of using this? Why is this important? Well, for example, we have our agents that we use. So we've got Claude, for example, we've got Hermes agent, we've got OpenCode, and right now Claude Opus is frontier level. And you've got ChatGPT as well, which is frontier level. So you've got two main frontier models who are absolutely dominating in AI right now. And so with this model, it's actually a lot faster than something like Opus 5.5.

But at the same time, it can perform on the same level. So, for example, we can plug it into our Agentic Harness, or we can use it directly inside something like OpenCode. Now, OpenCode just dropped DeepSeek V4.1-Flash-3, and then you've got the newest version, which is this one right here. So let's say, for example, we could ask it to build out a 3D game or something like this. And the cool thing about this is when you're building out, it can use skills and it can operate in the same way that Opus 5 can operate, but at the same time is way, way faster and it's also a lot cheaper. So you might be wondering, okay, why is this worth understanding? Why is this worth really taking a look at? So it's a Flash model. A Flash model is trading blows with the frontier. Flash is DeepSeek's cheap, fast tier and on DeepSeek's own table, it actually beats Claude Opus 5 on four of 11 benchmarks. All of them are Gentic or Koji, right? And budget models, they're not really supposed to do that. That's the main thing here. Now, also, the gains, the improvements, they come from training, not size. So it's literally like, and DeepSeek always do this, like the same architecture, same parameter count, and every single benchmark just went up by two to 11 points just in the post training, which makes you think like post training makes a huge difference for AI model. It's also built for AI agents and price for them. So agents use up a lot of tokens, especially in long loops. And so model that's strong at multi-step tool use finishes with shorter outputs and costs way less. And also DeepSeek have opened its own web search. It's pretty interesting. So on the same day, it's search and fetch APIs went public as well. And it's free through DeepSeek Harness, Claude Code or Codex until the 1st of November, 2026

So they've basically added like a web search layer directly into DeepSeek Harness as well. And you can use that elsewhere too. So let's get into the benchmarks here and see how it performs versus other models if we open this up. So the first one we will look at is can it write and fix real software? And these are three benchmarks, three different jobs, DeepSWE 1.1 is the main one. And in this test, basically the model gets a real code base and a real bug report and has to ship a fix that passes the project's tests. This is the closest thing here to a day of software engineering work. Now, if we look at the benchmarks, top of the table, 1017 beats Opus 5 by 3 points and GPC 5.6 Sol by 4 points. The clearest coding win in the whole release. That is absolutely massive, right? V4.1-Flash is beating Opus 5 Now, something to note here is you don't see GPC 6 in the comparisons. You're not seeing Opus 5.5 in the comparisons. But still, these are friendly front and GPC V4.1-Flash is actually beating.