Core Concepts at a Glance
- Open weights, open reasoning. Because the model is open-source and open-weights, Berman can see every chain of thought and tool call, and can download and run it himself.
- Flash means fast and cheap. V4.1 Flash is described as a fraction of the price of the top closed models and very fast, with provider speeds above 800 tokens per second.
- Multimodal. The model accepts images, and Berman tests it on a hard counting task with hidden objects.
- Simple prompts as a test. Berman deliberately uses short prompts (a Fall Guys clone, a Rubik's cube) to see whether the model builds what a person already has in mind.
- Cost per intelligence task. A metric that combines a model's price with how many tokens it needs to finish a task.
- Pareto frontier. On a chart of quality against cost, the models that are not beaten on both at once; Berman says DeepSeek sits on it.
- Build your own software. A model this cheap makes custom apps (such as his PDF reader, Folio) practical.
Introduction
In this 21-minute video, Matthew Berman argues that DeepSeek's latest model, DeepSeek V4.1 Flash, flew under the radar and is the best open-source model available. He supports the claim with a series of live demos (3D builds, image understanding, speed, browser use, and apps built locally) and then with benchmark and pricing charts. His final position is careful: it is not the highest-quality model, but with cost, quality, and speed combined he thinks it is number one. Figures below are as stated in the video, and the auto-captions garble several names, so verify specifics before relying on them.
1. Hands-On 3D Builds (00:00)
What it is
Berman opens with tests that combine 3D generation, spatial awareness, and knowledge stored in the model's weights.
- Mechanical watch (00:20). The prompt asks for a hyperrealistic 3D mechanical watch diagram with hover labels for each part. The result, which he calls Mechanica, is a single web page with working lighting and moving shadows, hour, minute, and second hands, and parts that fit together. Clicking a part such as the hairspring highlights it, centers it, and shows a description; an examine option zooms in, and an explode option separates each layer while hover still works (02:01). His one complaint is that the weight on the back that charges the watch from wrist movement is missing (01:43).
- Fall Guys clone (02:28). A one-line prompt (build a Fall Guys clone) produced a 3D game called Bean Blitz in HTML. Hammers knock the player off, dive works with the space bar, and the 3D looks good. He notes some object clipping and a spinner that does not actually spin. He says it is not as good as Opus 5.5 but quite close, and the model is a fraction of the price and very fast (03:34).
- iPhone Duo (03:49). This test requires the model to research the web for the device's specs and design and rebuild it in 3D. Berman calls the result flawless (logo, light reflection in the screen, matte inner screen, buttons, hinge). He ran the same test on the new Kimi model and says it shows DeepSeek is much better (04:28). The fold animation has a flaw: the device drifts off screen and the screen briefly disappears while folding.
Why it matters
Strong results on a single-prompt, single-file build show real spatial reasoning and world knowledge, not just code completion.
How to apply
Test a model with one-line prompts for things you already know well. If it builds what you pictured without extra detail, it is carrying that knowledge itself.
2. Sponsor: CoreWeave Arena (05:00)
What it is
The sponsor segment describes CoreWeave Arena, a way to load test an AI app before it ships.
- What the demo shows (05:12). In Forge, CoreWeave's AI development platform, requests are sent one at a time, then five, then ten, while response time, throughput, and estimated cost are tracked. When ten requests took longer than five, an assistant called Arya recommended reducing unnecessary processing for each answer. Arena is described as a guided program for testing, working with CoreWeave's engineers.
Why it matters
Berman ties it to using this model in a production app: check it before launch.
How to apply
Before putting any model behind a product, test it under realistic concurrent load and watch latency, throughput, and cost.
3. Visual Understanding: Counting Pastries (06:05)
What it is
V4.1 Flash is multimodal, so Berman gives it a photo of a pastry counter and asks it to count each item.
- First count (06:24). After 4 minutes 25 seconds it reported 137 pastries and named each type. For the macaron pyramid it counted 24; Berman counts about 25 visible and calls 24 arguable. It listed eight butter croissants and nine cream-filled horned croissants (07:15).
- Unprompted caution (07:31). Without being asked, it noted that macarons, croissants, and brownies are stacked so some are hidden, and gave a defensible range of roughly 130 to 145.
- Estimating the unseen (07:51). Asked to include pastries it could not see, it estimated 21 more hidden macarons, about 45 in total, reasoning that the heap looked like a cone five to six pieces wide and five to six layers tall (08:27). For the croissant piles it estimated two to three layers deep and added about 60% to the visible count (08:49).
- Timing (08:55). The original count took about 4.5 minutes; the follow-up took another 23 seconds because the work was already done. This was on max reasoning.
Why it matters
The test goes beyond naming objects: the model reasons about what is probably there but out of view.
How to apply
When you use vision models for counting or inventory, ask for visible and estimated-hidden counts separately, and check the visible count yourself, as Berman does.
4. Speed Test and Inference Providers (09:12)
What it is
The Flash name signals speed, and Berman tests it with reasoning off and a prompt to write a 1,000-word story directly in the chat without using tools.
- Why no tools (09:40). Even with thinking off, a model may decide to create a document, so he tells it not to use tools and to output the text in the window.
- Result (10:04). He estimates 125 to 150 tokens per second in that setup, and notes this is not the fastest available.
- Providers (10:10). Citing Artificial Analysis, he says Lithos AI serves the model at up to 1,286 tokens per second and Base 10 at 845.
- His conclusion (10:27). Not the highest quality, but close, and with cost, quality, and speed together he thinks it is the number one model overall.
Why it matters
Speed matters for interactive use and for long agent runs, and open weights let several providers compete on it.
How to apply
Compare providers for the same open model on speed and price, since they can differ a lot.
5. Browser Use: Finding Clippable Moments (10:41)
What it is
This test checks four things: using a browser, understanding YouTube, recognizing a clippable moment in an interview, and giving correct timestamps.
- The task (11:01). Berman gave it his Sam Altman interview and asked for the top five most clippable moments with exact timestamps and reasons.
- The run (11:12). It worked for 6 minutes 24 seconds. With an open-weights model he can read the whole chain of thought and every tool call.
- The clips (11:21). The first clip was a quote about not having looked at a line of code in about a year; Berman verified the timestamp was correct. The model explained why it would work as a short (newsworthy, vivid, self-deprecating, no setup needed). The others included that coders have never been busier and that the consciousness debate can be a backdoor to taking power from people (12:15). He says he would have chosen the same clips.
- Next step (12:35). He has not tried it yet, but wants to see whether the model can download the video and cut the clips itself.
Why it matters
It combines tool use, long-video understanding, and editorial judgment, and the timestamps were accurate.
How to apply
Ask for exact timestamps and the reason for each pick, then spot-check a few against the video.
6. Running Locally: Excel Clone and Rubik's Cube (12:48)
What it is
Berman ran the model locally on a DGX Station, which he acknowledges most people do not have, and says quantized versions can run on a high-end consumer GPU (13:06).
- Excel clone (13:23). The prompt was to build an Excel clone for the browser. After about 50 minutes it produced GridLine, an offline-first spreadsheet. It lacks full Excel functionality but handles formulas (a cell equal to B7 plus C7 computes correctly), bold, dragging a formula down a column, and inserting charts.
- Rubik's cube (14:09). A deliberately simple prompt: build a 3D simulation of a Rubik's cube, scramble and solve. Colors stay on the correct sides, the scramble moves are shown, rotations and physics look accurate, and the solve works (15:00).
Why it matters
A capable open-weights model you can download and use freely changes who can run strong AI, including offline and private work.
How to apply
If you have the hardware, try a quantized local version for long builds, which avoids per-token cost entirely.
7. Build Your Own Software: Folio (15:09)
What it is
Berman argues you should be building your own software when existing apps are missing, poor, or bloated.
- The example (15:28). He wanted a Windows PDF reader without Adobe Acrobat's bloat. He first built it with Opus 5.5, then had Opus 5.5 write a very detailed prompt and pasted that into DeepSeek on high thinking.
- The prompt (15:57). It asked for a PDF reader and annotation app called Folio using Electron, React, TypeScript, and PDF.js, with a warm off-white light theme, a near-black dark theme, subtle borders, compact controls, drag and drop, and text highlighting.
- Features (16:28). Pages and table of contents, zoom, drawing in different colors, movable drawings, text notes, and signatures created by drawing, uploading, or photographing a signature on paper. It also has a dark mode for any PDF, export, rotate, and print.
- Cost (17:28). He ran it locally, and says through an inference provider it would likely cost less than a couple of dollars.
Why it matters
Cheap, capable models turn a custom tool from a project into an afternoon.
How to apply
Let a strong model write a detailed spec prompt, then hand that prompt to a cheaper model to build.
8. Benchmarks, Cost per Task, and Pricing (17:45)
What it is
Berman closes the evidence with charts, saying benchmarks only tell part of the story after the demos.
- Quality versus cost (18:14). On a quality-and-cost chart he says DeepSeek sits on the Pareto frontier and beats every other model shown, which includes Opus 5, Fable 5, Gemini 3.8 Flash, and GLM 5.3 Max. Up is better and left is cheaper, and he calls it the best bang for the buck unless you are doing the most complex tasks.
- Other benchmarks (19:02). Humanity's Last Exam at 42, on par with Kimi K3 and GLM 5.3, and Terminal Bench at 42.
- Cost per intelligence task (19:24). This combines a model's price with the tokens it needs for a task. He reads out the top figure for Claude Fable 5.1 Max at $763 and Kimi K3.1 Pro at $211, against 26 cents per completed task for DeepSeek (20:15).
- Pricing (20:21). On Lithos AI he reads 60 cents per million output tokens and 15 cents per million input, with cache hits free. On Base 10 he reads 30 cents per million input and one cent per million cache-hit tokens, with the output price garbled in the captions.
- Kimi K3.1 (20:56). He notes Moonshot's newest model scored higher overall on benchmarks but personally thinks DeepSeek is much better, and tells viewers to compare for themselves.
Why it matters
Price, tokens used per task, and speed together decide real-world value more than a single benchmark score.
How to apply
Judge a model on cost per completed task and on your own tests, not on sticker price per token or a headline benchmark.
Closing
Berman's case is that DeepSeek V4.1 Flash delivers near-frontier results at low cost, high speed, and with open weights, and that this combination makes it the best overall model right now, though not the highest quality. The demos back up the practical side (3D builds, vision, browser use, local app building), and the charts support the cost side. Treat the rankings and prices as his readings from the moment of recording, and run your own comparison against Kimi K3.1 and your current model before switching.