DeepSeek V4.1 Flash 1017 Is INSANE! Best Open Model EVER?! (Fully Tested)

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: DeepSeek V4.1 Flash 1017 Is INSANE! Best Open Model EVER?! (Fully Tested) · Channel: WorldofAI · 20:40 · Watch on YouTube

Topics: AI agents, AI evaluation and benchmarks, Software engineering with AI, AI models, Local and open-source models

Summary

The video reviews DeepSeek V4.1 Flash 1017, an updated version of the Flash model that DeepSeek says improves performance, speed, and output length without increasing cost. The presenter highlights its low API pricing, large context window, agentic coding abilities, and strong results in front-end, 3D, and other coding tasks, while noting that some projects take much longer than competing models. Comparisons show that it can approach more expensive models on selected tasks, though quality and speed vary. The presenter recommends trying it, while emphasizing that it is not perfect and that claims about distillation from OpenAI are speculative.

Key points

V4.1 Flash 1017 is an update to DeepSeek V4.1 Flash, rather than a new flagship model.
The video cites pricing of $0.30 per million input tokens and $1.20 per million output tokens, with a one-million-token context window.
DeepSeek reports improvements on Terminal Bench 4.0 and HLE; the presenter also says the model ranked around 10th or 11th in the World of AI benchmark.
The model is praised for agentic coding and for generating front-end projects, 3D scenes, games, and other complex outputs.
DeepSeek’s search and fetch APIs are described as costing $0.75 per 1,000 calls, with server-side tool calls free through October 31 at 16:00 UTC, as stated in the video.
The presenter recommends using DeepSeek Harness to help the model test, identify errors, and refine its work.
In several comparisons, DeepSeek produced results at substantially lower cost than competing models, but often took longer or delivered lower overall quality.
The presenter says the model is too large to run locally in practical settings.
The suggestion that DeepSeek may have distilled OpenAI models is presented as speculation, not confirmed evidence.

Chapters

Start	Chapter	Summary
00:00	What DeepSeek V4.1 Flash 1017 offers	The presenter introduces the update, its availability, focus on agentic development, and DeepSeek’s search and fetch APIs.
03:04	Benchmarks, pricing, and speed	The video discusses reported benchmark gains, the model’s price and context window, and its reported inference speed.
06:08	3D and front-end generation tests	The presenter compares older and newer DeepSeek outputs and showcases front-end, product-viewer, and SVG tasks.
09:14	Comparisons on creative coding tasks	The video compares DeepSeek with other models on animations and voxel art, focusing on quality, completion time, and cost.
12:23	Games and the DeepSeek Harness	The presenter compares parkour game outputs and explains how a harness can support iterative tool use and code refinement.
14:26	Further model comparisons	The video compares DeepSeek with GPT-6.1 Soul Ultra and Kimi K3 on complex tasks, noting tradeoffs in cost, time, and output quality.
17:33	More 3D tasks and final assessment	The presenter reviews an SVG comparison, Blender work, and video creation before summarizing the model’s strengths and limitations.
20:08	Conclusion and sign-off	The presenter recommends trying the model, reiterates that it is imperfect, and closes with channel links and sign-off.

Quotes

they’re calling this a Pareto improvement — 00:00

30 cents per million input tokens and $1.20 per million output tokens — 04:05

the model can continuously test, identify errors, and refine its outputs — 13:55

the price to performance is definitely insane with this model — 19:37

that is just something that every model provider does. So, don’t quote me on that. — 20:08

Concepts and entities

DeepSeek V4.1 Flash 1017, DeepSeek, DeepSeek API, DeepSeek Harness, Terminal-Bench 4.0, HLE, OpenAI, Hermes Agent, Blender, WorldofAI

Full transcript (original language)

00:00 This was completely unexpected because after a long slumber, the whale is finally back and DeepSeek is looking to make another serious punch in the AI space with the release of DeepSeek version 4.1 flash 1017. And this update is actually quite impressive cuz DeepSeek is claiming stronger benchmark performance, faster task completion, and shorter outputs compared to the previous versions. Essentially, they’re calling this a Pareto improvement, meaning better performance at the same or lower

00:30 cost. Now, right off the bat, something worth mentioning is that this isn’t an entirely new flagship model. It is an upgraded version of DeepSeek version 4.1 flash, which originally launched on September 10th. And DeepSeek has continued refining the flash series over the past couple of months with multiple updates, with the latest one being released just yesterday. And this model is currently available through DeepSeek, which is currently being routed under DeepThink as well as their regular chat.

01:01 And you can even access it through the mobile app as well as the API, which you can access starting today. What’s particularly interesting is that DeepSeek has been focusing heavily on agentic coding and project development. They’re using reinforcement learning with self-feedback to improve the model’s ability to complete entire projects, whether that’s generating games, building apps, or even working on 3D modeling tasks. And that’s a pretty important distinction because we’re beyond just models that can simply generate basic code towards models that

01:32 can actually complete complex development workflows from start to finish. DeepSeek is also making another interesting move by opening up its search and fetch APIs to everyone. Sorry guys for interrupting, but if you guys haven’t already, please go ahead and like this video and subscribe as majority of you guys aren’t actually subscribed. And it truly helps me make videos for you guys on a day-to-day basis. So, I’d really appreciate it if you guys can go ahead and like this video. It would really mean a lot to me. These are powered by DeepSeek’s own search infrastructure, which the company

02:02 says that is going to be able to handle hundreds of millions of requests daily worldwide. Both endpoints cost just 75 cents per 1,000 calls, which is quite competitive. And here’s the interesting part. If you’re using the DeepSeek API with tools like Cloud Code, Codex, or DeepSeek Harness, which is the best harness I think you should be using this model with, search and fetch calls through their server-side tools are currently free until October 31st at 16 UTC. So, essentially, DeepSeek is not

02:33 only improving its coding models, but also building up tools that are needed for making this model more capable as an autonomous agent. Of course, independent evaluation of this specific model is still quite limited, which is why we’re going to be testing out this model throughout today’s video. But, considering how aggressive DeepSeek has competed with its price, sync, efficiency, and open-weight models, it’s definitely exciting to see them back again making this new update. Now, looking at the benchmarks, DeepSeek has

03:04 made some pretty solid improvements from its previous model, cuz on Terminal Bench 4.0, the performance jumped from surpassing GPT 5.6 Soul at 39.9% on DeepSave version 1.1. It scored a 77%, which is beating GPT 5.6 Soul, Claude Opus 5, and GLM 5.3 in DeepSeek’s report comparisons. On HLE, which is where it is recording an improved score

03:34 of 42.6% from a previous 36.8% score, showcasing this noticeable jump in reasoning as well. And it overall just does really well in terms of completing tasks faster with shorter outputs with the same low pricing. And we have tested out this model rigorously through the world of AI benchmark across multiple domains. And what we have found is that this model currently is ranked number 10th overall, which is or 11th overall, which is quite

04:05 impressive. It’s obviously not hitting the top 10 just yet cuz it is slightly behind in a couple of areas, but it is quite impressive with its agentic coding capabilities. And in regards to pricing, this is a model that is listed at 30 cents per million input tokens and $1.20 per million output tokens with a 1 million token context window. This currently makes it the best open-source model. Keep in mind, this is a model that isn’t practically locally available

04:36 as it is just way too massive for anyone to actually run locally. Also, if you want to run benchmarks on your own model as well as unreleased models as well as getting access to those unreleased models and even accessing free models, you should 100% take a look at the world of AI benchmark cuz there are a lot of cool things that you can do with it like building out your own product, even exporting those comparison videos where you test out multiple models, even getting AI leaks where you get up-to-date leaks every single hour of the day, and a lot of other things. So,

05:08 I highly recommend checking it out so that you can get started and learn which model you should be using for your app. Now, I didn’t really want to rush into this model right away, which is why I’ve been testing it a lot over the past couple of hours, and this is a model that I have found to be insanely fast while also delivering some crazy results. Cuz it is something that reportedly hit about 244 tokens per second with some early tests nearing 300 tokens per second, which is insane. And

05:38 that’s potentially even faster than GPT-6.1 Soul in ultra-fast mode while delivering intelligence that is comparable to GPT-5.6 Soul. Even the benchmarks are showing that it is surpassing it in many areas, which is kind of ridiculous that this cheap model is able to be this capable. Deep Seek has also introduced this built-in web research capability through its search and fetch tool. So, that is something that is free and I would highly recommend making the most out of it as it was able to get me better outputs.

06:08 And another tip I would recommend is using it with the harness, which is going to let you get more out of this model. If you want the best AI tools, workflows, and drops before everyone else, join my free newsletter with the link in the description below, which is completely free. To start things off, let’s take a look at a couple of these 3D scenes that the Deep Seek models were able to generate. This is a comparison test with the original Deep Seek version 4.1 flash, the original one that is 2 months old versus what the new one had

06:39 generated. And there is a pretty big difference that is night and day. The newer model delivers a noticeable better a noticeably better result, showing just how much Deep Seek has improved its 3D generation capabilities with a single update. One thing I’ve really love is that this model has gotten insanely good at front-end development. It is something that has taken a pretty big leap in this specific domain over a lot of different sorts of generation styles. This is something that it did quite well

07:10 with a couple of the front-end tests that I’m going to be showcasing throughout today’s video. And you can see that with the typography, the dynamic movements, the small little subtle effects that you can see with these small little animations, as well as the scroll triggers, everything seems to be fluid. And I just love how it’s able to output this sort of quality with the pricing that we have listed for this model. Here is the 360 product viewer, and it is pretty solid. Now, one thing you will notice right away is that the output quality kind of

07:42 mimics what Open AI is generating. Not saying that they are 100% copying Open AI, but I have some information about that, which we’re going to be talking about. But, still, regardless, I am happy that we’re getting this sort of output where it fully coded this 3D visual of a product, which is something that most models actually tend to fail at doing. But, in this case, the DeepSeek model was able to output it fluidly. And let me just showcase what I mean about the Open AI style outputs. So, apparently, this model has been

08:14 identifying itself as an Open AI model in certain responses when it is working on in-depth reasoning, which has sparked a lot of speculation that distillation may have played a role in its training. Of course, this isn’t definite proof that they are doing this, but it is very likely that DeepSeek’s team has been distilling the Open AI models, and you can see that with a couple of the different outputs that we’re going to be showcasing throughout today’s video. Now, in this specific prompt, the pelican riding a bike SVG prompt is

08:44 where the DeepSeek version 4.1 Flash has done quite well in this output. The visual output looks pretty impressive in comparison to the other models that you see over here, like Opus 1.5, Gemini for Argon, as well as GPT 6.1 Soul, which was set on X High. All of these models were set on the highest reasoning level, but you can see that the Opus model did obviously the best when it comes to different visual style. You can see that the DeepSeek model does pretty decent

09:14 with its overall output quality. Obviously, it doesn’t tell you the whole story that this model is better, but the price to performance DeepSeek is offering here is definitely a steal. Here, DeepSeek version 4.1 Flash was put up against the Opus 1.5 using the exact same prompt, and Opus obviously cooked heavily with this output, delivering better animations, sound, attention to detail. But, the interesting part is that DeepSeek did a pretty decent job with its overall output. Obviously, not the same level of quality, but it is

09:45 something that got the base structure down of whatever was requested and costed just 23 cents compared to Opus costing $6. That is making it roughly 26 times cheaper with the DeepSeek output. But, DeepSeek did take an hour to finish this, whereas with Opus, it completed the task in 20 minutes. So, DeepSeek’s output quality had a few errors, but it still managed to get surprisingly close in quality while being dramatically cheaper. Here is a voxel art generation,

10:17 and this is a great benchmark like I had always mentioned in my videos because it tests the AI’s model on reasoning with spatial understanding, coding ability, 3D understanding, and ability to maintain consistency across massive complex environments. And in this case, this is truly a massive environment cuz you can see that there’s multiple aspects to this medieval kingdom that was generated with the DeepSeek model, where it generated over 2 million voxels in a single file, and it completed the task in roughly 25 minutes, which is

10:49 quite fast for a voxel art generation with a cost of just $5 through the API. It is something that was able to add in its creative touch. It was able to also fluently edit out anything that didn’t actually incorporate itself into this generation, but it did quite well with the overall scene, where you can see there’s a siege of this medieval town, and it is being executed with all these other side panel features, like a day and night cycle, as well as adding in

11:21 other features, like adding in a dragon, which is just really cool, and it’s just nice to see that there’s this level of detail that is being outputted from this model. Many people have been recently comparing this version 4.1 flash model to the GPT 6 Luna, which I do believe is a pretty good comp, but it is slightly behind in many cases, but still the quality as well as the pricing does kind of mimic what Luna does. As this is one test that gives you that outlook where it was

11:53 comparing the DeepSeek version 4.1 flash with GPT 6 Luna, and both tasks or both models were tasked to create the Tesla Roadster event in a 3D environment. DeepSeek completed this task for about 89 cents compared to Luna’s $1.07 output, which is making it about 17% cheaper. Obviously, it varies for the task, but I noticed that the GPT 6 Luna output was a bit better in terms of its quality. If you’d like this video and would love to support the channel, you

12:23 can consider donating to my channel through the Super Thanks option below. Or you can consider joining our private Discord where you can access multiple subscriptions to different AI tools for free on a monthly basis, plus daily AI news and exclusive content, plus a lot more. Here is another interesting comparison where four flash models were tasked to build a first-person parkour game in a single HTML file. This features four stages and one round of fixes with DeepSeek version 4.1 flash reportedly

12:54 delivering the best visual effects. However, I find that the Haiku model did pretty decent job with the output quality, which made a more functional app as well, but I don’t know. It is quite debatable. DeepSeek’s generation quality does feature some better visuals in a couple of areas, but overall DeepSeek version 4.1 flash reportedly delivered through the output at a cost of 14 cents, which is pretty good, but it is something that

13:24 took about 1 hour and 30 minutes to generate this output, whereas something like Haiku only spent 8 cents, which is is drastically cheaper than DeepSeek, and it is something that did it in a much shorter time span. Something worth mentioning is like I had talked about that the harness that DeepSeek has works really well with this model, as it is something that is able to extract the capabilities of this model further cuz it’s a lethal combination because the harness is specifically designed around

13:55 DeepSeek’s models, cuz it optimizes how they use tools, manage contexts, as well as executing code, and work through complex tasks. Instead of just generating code in a single shot through the API, which could be fine in most cases, but in this particular case where you have it attached to a harness, the model can continuously test, identify errors, and refine its outputs thanks to the harness itself. So, when you combine the two, you’re going to be able to get improved agentic capabilities, where it’s going to be able to make the most

14:26 out of the model. This model is just fascinating. This is a test this time between GPT-6.1 Soul Ultra up against DeepSeek version 4.1 Flash, and this is with both models running on max reasoning. DeepSeek was reportedly about 90% cheaper, just costing 44 cents through the API, yet managed to deliver some pretty impressive results. Both models took roughly the same amount of time, although most of that was spent in rendering the outputs out. But, what’s

14:57 interesting here is that DeepSeek was able to access the web. It was able to use those tools, which is a significant advantage when gathering references and building complex projects out. And it was able to get this pretty decent output thanks to that feature. And remember, this is something that is free. The web searching API inputs are currently free right now with DeepSeek, so it is something that you should make the most out of. and considering that this is a model that also has multimodal capabilities with this ridiculously low

15:27 pricing, the value is truly here, especially if you’re getting near Soul outputs. When you are to take a look at something like this, it makes you think that it was outputted by a model like Opus, but this is actually DeepSeek version 4.1 Flash, where it was requested to create a realistic grassy hill with detailed grass, a tree, as well as dandelions blowing in the wind, and it managed to generate the entire scene in just one turn. But, what’s even more impressive is the attention to detail with this output, from the wind

15:59 effects to the realistic environment, and it even generated sounds that go along with the scenes. And remember, this isn’t some flagship model like Opus. This isn’t open-source model, it’s a Flash variant as well, and it was able to output this sort of quality while being pretty cheap, and that is just a steal in my book. Here we have the Minecraft prompt with Kimi K3 put up against DeepSeek version 4.1 Flash, where both of these models were tasked to build this sandbox game, and we can see that both of them did a

16:30 pretty decent job. Both of them used their highest reasoning effort, and one thing I can notice right away is the actual depth that Kimi was able to output is pretty impressive. It even added in another dimension from the original game, and that is actually pretty cool and how it’s able to use these aspects to output these sorts of generations, but Kimi did take 2.5 hours and cost $35, which is definitely expensive and quite long to get that sort of output. But, it managed to build

17:02 a lot of these cool dimensions out, which is something that was able to even output some pretty impressive graphics. Meanwhile, with DeepSeek, it took about 1.5 hours, and it cost about 80 cents, which is nearly 44 times cheaper, but in many cases it did struggle with a lot of the features, which is nowhere near in terms of the same quality as Gemini K3. However, with multiple prompting attempts, it is something that can definitely get there. Back to SVG, this

17:33 is a test with Gemini 4 Argon up against the DeepSeek version 4.1 Flash, where both were tested on the exact same prompt at their highest reasoning effort. This is where Argon finished this task in about 13 minutes for about $1.20, but DeepSeek took about 50 minutes, which is definitely quite long for an SVG output, but it only cost 40 cents, making it about three times cheaper. But what’s surprising is how close the final results were, with DeepSeek reportedly delivering even better water detail in

18:03 certain areas, but Gemini 4 Argon did generate the better overall scene while completing the task faster. So, it’s kind of weird. Sometimes the model does quite well with its outputs, but sometimes it is a little slower. Based off of some of these tests and running my own tests, I can clearly tell you that this model has gotten a lot better at 3D outputs. Maybe because it has been distilling from OpenAI, but I will tell you that it is doing these tasks at cheaper price, but it may be

18:34 longer in getting those outputs. Now, this is actually pretty cool cuz DeepSeek version 4.1 in Hermes Agent, where it used Blender, and it is something that took about 7 hours, which is a lot of time, but to be fair, this is a huge improvement from the previous generation, where it was also taking a longer time and it spent a lot of time. But in this case, it is something that generated this 3D project using these tool assets, and it was able to do this at a cost of just $1.83,

19:06 which is pretty remarkable. So, maybe if you’re going to be using this model overnight as you sleep, that could be something that is going to let you handle these tasks which may take a longer time to get the outputs that you’re looking for. And surprisingly, this model is actually pretty good at video creation. The only hindrance is the time spent in terms of creating these assets. But, this is a really, really nice output from this model that is showcasing Cerebras’ new product. It goes into detail and it showcases all

19:37 those things through these visualizations, which is really nice to see. Overall, I think DeepSeek version 4.1 flash 1017 is genuinely one of the most impressive flash models I have seen in quite some time. The coding capabilities, the 3D generation, and the price to performance is definitely insane with this model. So, it is something that I highly recommend checking out. You can try it out for free right now through their chat, but it is something that is still not perfect. There is a lot of things that could be better. I do find it to be

20:08 really close and similar to OpenAI’s models due to the distillation, but that is just something that every model provider does. So, don’t quote me on that. But, the fact that it’s able to compete with much more expensive frontier models is really nice to see and the fact that DeepSeek is open sourcing this model is a great benefit to the community. So, I highly recommend checking it out with the links in the description below. Huge thanks to DeepSeek for creating such a great model and allowing it to be something that most people can access for free. I’ll

20:40 leave all the links that I used in today’s video in the description below like the World of AI’s cool community as well as the World of AI benchmark. Take a look at the second channel. Join us on the newsletter, Discord, follow me on Twitter. And lastly, make sure you guys subscribe, turn on the notification bell, like this video, and please take a look at our previous videos so that you can stay up-to-date with the latest AI news. But, with that thought guys, have an amazing day. Spread positivity and I’ll see you guys fairly shortly. Peace out, folks.