Matthew Berman Tests DeepSeek's Latest Model Across 3D Design, Coding, Vision, Speed and Cost

Berman opens by calling DeepSeek's latest model the best open-source model on the planet, saying it had flown under the radar. His first test combined several capabilities: generating 3D models, spatial awareness, and knowledge stored in the model's weights. He asked for a hyperrealistic, beautiful 3D mechanical watch diagram in which hovering over each part shows its name and function.

According to Berman, the output fit together correctly, with lighting and moving shadows visible as the watch face was moved. Hovering over the hair spring displayed its name and function, and clicking it zoomed in and highlighted the part. The video's on-screen description calls the hair spring a flat spiral of hardened steel, thinner than a human hair, that stores and returns the balance wheel's kinetic energy. Each system and the overall parts could be clicked, and an explode view separated all layers while still allowing hover.

Berman's only criticism was that the back of the watch was empty. He said it was missing the rotor weight that moves with the wrist to charge the watch, though he said he could not recall the component's name at that moment. He described the rest of the result as incredible.

Next, Berman asked for a Fall Guys clone using only the prompt 'Build a Fall Guys clone.' The result, named Bean Blitz, was built as a single HTML web page. He said the 3D generation looked good, though some objects showed clipping. Obstacles such as hammers knocked the player off, and diving worked with the space bar. Some elements, such as spinning objects, did not actually move.

Berman said he had previously run this test on other models and that he would compare the results. His verdict was that DeepSeek's ability to clone the game was impressive. He said it was not as good as Opus 5.5 but was 'quite close,' and added that the model costs a fraction of the price and is very fast.

Berman then asked the model to create a 3D model of the iPhone Fold. He explained that this required it to search the web, understand the specification and design, and recreate the device in 3D space. He said it 'absolutely crushed this test.' He judged the Apple logo, the light reflection on the screen, the matte finish of the inner screen, the button placement and the hinge to be correct, calling the result flawless.

Berman said he ran the same test with the new Kimi model and that the comparison showed DeepSeek is much better. The on-screen comparison is referenced in the video, but the details of Kimi's output are not described in the transcript.

Berman then folded the phone in the simulation. He noted a small issue: when the folding was triggered, the animation partly went off screen. He said this was a problem with the simulation rather than the model's design. Once folded, the screen saver briefly disappeared during the fold, but he said overall the result looked incredible.

The video includes a sponsor segment for CoreWeave Arena, a program in which users can load test an AI app before production. Berman showed a load test that started with one request, then five, then ten, measuring response time, throughput and estimated cost, and said the assistant recommended reducing unnecessary processing for each answer. The sponsor segment is presented as an advertisement and is not part of the DeepSeek evaluation.

Berman uploaded a photograph of pastries, noting that DeepSeek V4.1 Flash is multimodal and accepts images. He described the counting as a very hard task. He counted the pastries himself, front to back, taking about four minutes and 25 seconds, and reported a tally of 137 distinct pastries visible on the counter.

For the orange macarons stacked in a pyramid, the model counted 24. Berman counted 25 visible macarons and said 24 was very close and arguable, noting that many macarons were hidden from view. The model also named the other items, such as eight butter croissants and nine cream-filled horned croissants, and, without being asked, noted that stacked macarons, croissants and brownies were partly or fully hidden. It gave a defensible range of roughly 130 to 145 pastries for the whole counter.

Berman then asked the model to estimate the total, including items that could not be seen directly. It estimated about 21 additional hidden macarons, reasoning that the heap reads as a cone roughly 5 to 6 pastries wide at the base and 5 to 6 layers tall, and that a triangular front face of 24 implies a base layer of roughly 19 to 25. It put the total at about 45 macarons, which Berman said was much more likely than 24.

For the butter croissants on the back stand, the model counted about seven and a half visible and estimated five more stacked behind them. For the horned croissant piles, it estimated 2 to 3 layers deep, adding about 60 percent to the visible count. Berman said he was 'beyond impressed,' and noted that the original count took about 4.5 minutes while the follow-up took only 23 more seconds because the calculation was already done, all on max reasoning.

Berman tested speed by selecting DeepSeek V4.1, turning off reasoning effort, and asking for a 1,000-word story in the chat window without tools. He said he disabled tools because, even with thinking off, the model might otherwise try to create a text document. The output appeared at what he estimated to be roughly 125 to 150 tokens per second.

Berman then referenced Artificial Analysis data showing that Lithos AI can serve the model at 1,286 tokens per second and Base 10 at 845 tokens per second. He argued that because of its cost, quality and speed, the model might be the current best overall model, though not the highest quality. Those speed figures come from the hosting providers and the chart he displayed.

Berman tested whether the model could use a browser, understand YouTube, judge which moments in an interview are clippable, and give correct timestamps. He gave it a Sam Altman interview and asked for the top five most clippable moments, with exact timestamps and reasons. The task took 6 minutes and 24 seconds. Because the model is open source with open weights, Berman said he could view its chain of thought and every tool call.

The first clip selected was Altman's remark that he had not looked at a line of code in about a year. Berman said the timestamp was correct and that he verified it. The quote given was: 'I realize I have not read or even looked at a line of code in probably a year. I don't think I'll ever do that again. Or if I do, it'll be like a Revolutionary War reenactment when I'm old and bored and trying to remember the glory days.' The model's reasoning, as Berman summarized it, was that an AI company CEO admitting this is newsworthy and the image is funny and widely understandable.

The third selection concerned the AI consciousness debate. Berman said it was one of his favorite parts of the interview. In it, Altman said: 'I get very nervous when people start using the AI consciousness debate to push into, well, maybe people shouldn't always be in control of the future. The future has to be about people.' Berman said he would have chosen these clips himself. He said his next step would be to have the model download and clip the video itself, but that this would be for another video.

Berman said he ran a quantized version of DeepSeek locally on a high-end consumer-grade GPU, while noting he owns a DGX Station, which most people do not have. Using the single request 'build an Excel clone for the browser,' the model worked for 50 minutes and produced GridLine, an offline-first Excel clone. Berman said it does not have full Excel functionality but is very good nonetheless.

In his demonstration, 23 plus 44 evaluated to 67, a cell formula referencing other cells worked correctly, bold formatting applied, dragging a formula down copied it into the other cells, and charts could be inserted and dragged. Berman then gave a simple prompt to build a 3D Rubik's Cube simulation that could scramble and solve. He said the colors stayed on the correct sides, the rotations and cube physics looked accurate, and it solved the cube.

Berman argues that people should now be building their own software. He says he has recently built several apps he needed because either they did not exist or the existing apps were poor. His example is a PDF reader for Windows. He said he did not want Adobe Acrobat, which he described as bloated and spammy, and said he did not need 90 percent of its features.

He said he had first built the reader with Opus 5.5, asking it for a detailed prompt, then pasted that prompt into DeepSeek's high-thinking setting to build a complete PDF reader and annotation app called Folio. He said this was a prompt very different from his usual tests.

Folio was specified to use Electron, React, TypeScript, and PDF.js for rendering. Berman said the prompt set the look and feel, including a warm off-white light theme, a near-black dark theme, subtle borders and compact controls, and described the desired PDF reading behavior such as drag and drop and text highlighting. He said the result was a beautiful, simple PDF reader and editor and that 'it just works.'

Berman listed features he demonstrated: a page and table-of-contents sidebar, zoom, drawing in different colors with movable drawings, text notes, and signatures created by drawing, uploading, or photographing a signature that is flipped and reused. It also offered dark mode, export, fit-to-page, clockwise rotation and printing. He said the app was basically free to make because he ran the model locally, and that if paying an inference provider, the cost would probably be less than a couple of dollars.

Berman says his enthusiasm rests on more than speed and low cost. He cites a quality-versus-cost chart, which shows DeepSeek V1.1 at the top left, where higher means better quality and further left means cheaper. He says the chart places it ahead of every other model shown, including Opus 5, GPT 6.1, Gemini 3.8 Flash and GLM 5.3 Max. He adds that the model gives up little quality unless the task is extremely complex.

Other benchmarks Berman showed included Humanity's Last Exam at 42, which he said is on par with Kimi K3 and GLM 5.3, and Terminal Bench at 42. He noted that benchmarks tell only part of the story.

Berman says that when evaluating a new model, the most important metric is cost per intelligence task, which combines a model's cost with the number of tokens it uses to complete a task. He showed figures including Claude Fable 5.1 Max at $763, Qwen 3.8 Max at $5, Astra at 3.26, and Kimi K3.1 Pro at $211, with DeepSeek at about 26 cents per completed task. He said he thinks DeepSeek is better than the newly released Kimi K3.1 Pro.

He also cited hosted prices. On Lithos AI, the model is listed at 60 cents per million output tokens and 15 cents per million input tokens, with cache hits free. On Base 10, he cited $1.20 per million output tokens, 30 cents per million input tokens and one cent per million cache-hit tokens, and said it serves about 800 tokens per second. He called the results impressive.

Berman closes by returning to Kimi K3.1, the newest model from Moonshot, which he said scored higher overall in the benchmarks. He said he personally thinks DeepSeek is much better and suggested viewers make the comparison themselves, referring them to his separate Kimi video.