Video summary
Deepseek crushed everyone
Main summary
Key takeaways
Summary
The presenter reviews DeepSeek V4.1 Flash, describing it as a fast, inexpensive open-weights model with strong coding, visual, and reasoning abilities. He argues that its combination of quality, cost, speed, and ability to run locally makes it one of the strongest all-around models available.
Demos and evaluations
- Interactive 3D mechanical watch: DeepSeek created a polished, browser-based 3D watch diagram with working-looking components, lighting, shadows, hover labels, explanations, zooming, and an exploded view. The presenter praised the detail and spatial layout, but noted that the back of the watch was missing its automatic winding rotor.
- Fall Guys-style game: It built a playable 3D browser game called Bean Blitz. The presenter found the visuals and basic mechanics impressive, including jumping, diving, and being knocked off by obstacles. Some objects appeared to animate without actually moving, and there was some clipping. He judged it close to, but not as good as, a version made by Opus 5.5.
- iPhone Duo model: After being asked to research and recreate the device, DeepSeek produced a detailed 3D model with accurate-looking buttons, hinge, logo, and finishes. The folding animation had an issue that caused the phone to move off-screen.
- Image understanding and counting: Given an image of a pastry counter, the model identified and counted visible pastries, then estimated how many were hidden behind stacks. The presenter praised its ability to name pastry types and reason about the arrangement. The initial count took about four and a half minutes; the follow-up hidden-item estimate took about 23 seconds. These counts were estimates, not independently established ground truth.
- Speed: In a direct text-generation test, the model produced a long story very quickly. The presenter cited reported serving speeds of up to 1,286 tokens per second from Lithos AI and about 845 tokens per second from Base10.
- YouTube research and clip selection: The model reviewed a linked Sam Altman interview and suggested five short-form clip moments, with timestamps, quotes, and reasons each might work. The presenter said the timestamps and selections he checked were accurate and compelling. The task took just over six minutes.
- Excel-style spreadsheet: Running DeepSeek locally, the presenter had it build GridLine, an offline-first browser spreadsheet. It supported basic formulas, filling formulas by dragging, and creating charts, though it did not have Excel’s full feature set. The build took about 50 minutes on his high-end workstation.
- Rubik’s Cube simulation: The model generated a 3D cube that could be scrambled and solved. The presenter said the colors, rotations, and solving behavior looked correct.
- PDF reader and annotation app: Using a detailed prompt originally prepared with Opus 5.5, DeepSeek built Folio, an Electron/React/TypeScript app using PDF.js. Features included page navigation, drawing and highlighting, text notes, signatures, dark mode, zoom, rotation, printing, and export. The presenter said it ran locally and cost little or nothing in inference, depending on how it was run.
Overall assessment and reported metrics
The presenter’s main conclusion is that DeepSeek’s appeal is not just its speed or low price, but its performance across a wide range of tasks. He described it as especially compelling for building custom software and for users who want an open-weights model they can run locally.
He also showed benchmark and cost comparisons, claiming that DeepSeek ranked especially well on a quality-versus-cost measure and performed competitively on several other tests, including Humanity’s Last Exam and Terminal Bench. He emphasized that benchmarks are only part of the picture and presented a “cost per intelligence task” comparison in which DeepSeek was reported at 26 cents per task.
Pricing examples he cited included $0.60 per million output tokens and $0.15 per million input tokens from Lithos AI, and $1.20 per million output tokens and $0.30 per million input tokens from Base10. These figures and rankings are the presenter’s reported comparisons.
A sponsored segment demonstrated CoreWeave Arena, a tool for load-testing AI applications before production. The presenter showed increasing request volumes and monitoring response time, throughput, and estimated cost, with an AI assistant recommending a way to reduce unnecessary processing.
Reviews, guides, and comparisons
- The video is primarily a hands-on review, testing DeepSeek through interactive demos rather than relying only on benchmark scores.
- It includes a practical example of using a detailed prompt to build a desktop PDF app, as well as examples of running the model locally.
- The presenter compares results with Opus 5.5 and Kimi K3.1, and recommends viewers make their own comparison. He says DeepSeek is better in his own experience, despite noting that Kimi scored higher overall on the benchmarks he referenced.
- The video also includes a brief CoreWeave Arena load-testing demonstration.
Main speakers and sources
- Matthew Berman — presenter and evaluator; delivers the commentary and demonstrations.
- DeepSeek V4.1 Flash — model being tested; produces the apps, simulations, image analysis, and video-clip suggestions.
- CoreWeave — sponsor featured in the Arena load-testing segment.
- External sources cited by the presenter — Artificial Analysis and the named inference providers Lithos AI and Base10, along with comparisons to Opus 5.5 and Kimi K3.1.
Rate this summary
Your feedback will help improve summaries.
Improve this summary
Reprocess with a stronger model when the summary feels incomplete or inaccurate.
Translate summary in another language
Ask questions to this video
Chat for follow-up questions, clarifications, and source-backed answers.