Matthew Berman calls DeepSeek V4.1 Flash 1017 one of the most compelling open-weight models he has tried. His video makes that case through a 3D watch, two game-like builds, image reasoning and personal software. The useful distinction for readers is what ran in DeepSeek’s hosted product, what Matthew ran locally, and what a new user can realistically run. Those are three different setups.
Try the updated model on DeepSeek’s web app first. For a coding agent or application, use the metered API model ID deepseek-flash. Download the official 1017 weights only if you have a serious self-hosting plan: the repository is about 510 GB, and DeepSeek’s reference runtime is not a one-click laptop install. The open-weight license makes experimentation possible; it does not make the hardware free.
Watch Matthew Berman’s Test
Source note, 28 October 2026: The builds, timings and qualitative comparisons below are Matthew’s observations in the supplied video transcript; we did not rerun them. Release status, API ID and prices, model-card benchmarks, license and repository size were checked against DeepSeek’s announcement, API documentation and its Hugging Face organization. The two links supplied with the request are official DeepSeek sources. Our setup and verification advice is editorial guidance, not an independently reproduced speed test.
What the Video Actually Tests
The video is strongest as a collection of working artifacts, not as a controlled league table. Matthew changes tasks, tools and reasoning settings; the competitor examples do not share a full published harness.
| Demo | What Matthew shows | What to verify yourself |
|---|---|---|
| Mechanical watch | Interactive exploded-view webpage with labeled parts and lighting. | Check mechanical accuracy, not just visual polish; Matthew notices a missing rear rotor. |
| Obstacle game | A playable Fall Guys-style browser scene. | Test collision, moving obstacles, win states and rights to copied design elements. He observes inert and clipping objects. |
| Foldable phone model | A polished interactive 3D product page. | Compare dimensions and controls against references; the fold animation briefly leaves the frame. |
| Pastry image | Visible-object count and a separate estimate of hidden pastries. | Keep observed counts separate from inferred hidden counts; the latter cannot be ground-truthed from one image. |
| Video clip research | Five interview moments with timestamps and reasons. | Open each timestamp and verify the quote before publishing a clip. |
| Local spreadsheet and PDF app | A spreadsheet clone built on Matthew’s high-end local machine, followed by a detailed-prompt PDF reader. | Test formula edge cases, exports, file security and document fidelity; these are prototypes, not full replacements for Excel or a PDF editor. |
What DeepSeek Confirms
DeepSeek released V4.1 Flash 1017 on 17 October 2026 for its website, app and API, and published the weights under MIT on its official Hugging Face account. It says the refresh improves reasoning, coding, agent tasks, visual benchmarks and task completion time against the September V4.1 Flash release. Its model card reports Terminal-Bench 4.0 rising from 31.2 to 42.3 and DeepSWE v1.1 from 74.2 to 77.0. These are vendor-reported, harness-dependent scores, not evidence that every 3D or coding task beats a specific Claude version.
The API supports text and image input, with a one-million-token context ceiling. A ceiling is not a recommendation to send an entire repository at once: large contexts still cost tokens, latency and attention. DeepSeek’s pricing page lists off-peak rates of $0.15 per million uncached input tokens and $0.60 per million output tokens, with peak rates of $0.30 and $1.20. Cached input has a separate lower rate, and peak windows are specified in UTC. Matthew’s third-party provider prices and tokens-per-second figures are snapshots from his recording, not DeepSeek’s published API guarantee.
Which Setup Makes Sense?
| Your goal | Start with | Trade-off |
|---|---|---|
| See whether you like the model | DeepSeek website or app | No local installation; account limits and hosted data processing still apply. |
| Build an agent or app | Official deepseek-flash API | Usage-based charges and tool costs; keep the API key private. |
| Control inference infrastructure | Official 1017 weights on a capable server | Roughly 510 GB download plus specialist memory, conversion and serving work. |
| Experiment on consumer hardware | Evaluate a third-party quantization separately | Different publisher, quality, speed, context and vision support; it is not the official checkpoint. |
Start Easily: Web or API
- Try the hosted app. Open chat.deepseek.com or the official app and confirm which model the interface offers. Give it one bounded task, such as a single interactive page, with success criteria. Do not upload confidential work just to test a demo.
- For API access, create an account and key in the DeepSeek platform. Read the first-call guide. Its base URL is
https://api.deepseek.comand the hosted model ID isdeepseek-flash, which DeepSeek says serves the 1017 release. Do not paste a key into source code or a public repository. - Keep the first spend small and test once. Top up only what you need for the experiment; use a short text prompt before adding image input, a browser, file access or an agent harness. Record total input, output, cache hits and wall time. The published peak/off-peak rates can change.
- Connect a coding tool only when needed. DeepSeek provides Codex integration instructions and guides for other harnesses. Read what any setup script changes before running it; preserve your existing configuration and use a disposable test project first.
For most readers, those four steps are the actual installation path: no model download, no GPU rental, and a clear bill. The API is a paid service even though the weights are openly licensed.
Advanced Route: Download the Official Weights
Only continue if the storage and hardware section below fits your machine. The official model card links its 48-shard checkpoint and a reference inference README. Downloading is not the same as running.
- Prepare an SSD and Python environment. Reserve at least 600 GB free for the original repository, more if keeping converted weights or caches. The exact download can be checked before it starts.
- Install the official Hugging Face CLI following its CLI guide, for example
python -m pip install -U huggingface_hub. Runhf --helpto confirm the command is present. - Preview the download:
hf download deepseek-ai/DeepSeek-V4.1-Flash-1017 --dry-run. Inspect the total, path and available disk space. Do not start a half-terabyte transfer on a small system drive. - Download to a chosen folder:
hf download deepseek-ai/DeepSeek-V4.1-Flash-1017 --local-dir /models/deepseek-v4.1-flash-1017. Replace the example path with an existing location on your large drive. On Windows, use a drive path such asD:\models\deepseek-v4.1-flash-1017. Keep the full 1017 revision together; do not mix shards from the September checkpoint. - Choose a supported runtime and follow its versioned instructions. DeepSeek’s own minimal inference path installs its requirements, converts weights into a tensor-parallel checkpoint (example
MP=8) and runs withtorchrun. It explicitly describes itself as a readable reference implementation, not a production server. Verify the runtime, kernel and GPU support before converting; conversion needs additional free space. Do not assume a generic one-line vLLM or Docker snippet on the Hugging Face page will work on your machine.
Storage, Memory and Cost Reality
Hugging Face displays roughly 510 GB for the official mixed-precision repository. Our 600 GB free-space suggestion is a planning buffer, not a published minimum. Conversion can require a second large copy; check the conversion output format and free space before starting. DeepSeek’s reference example uses eight tensor-parallel ranks, while its inference team discusses multi-GPU deployment. Neither source gives a trustworthy universal “minimum laptop GPU” specification.
A community quantization may fit a different machine, often by reducing weight precision or streaming from SSD. Treat each quantization as a separate artifact: check publisher, file size, license, checksum, actual image-input support, context limit and speed. Fitting the file on disk is not proof that it loads into memory or runs usefully. Matthew says his local spreadsheet build used an unusually expensive DGX-class workstation; his result should not be presented as a standard consumer-laptop benchmark.
One Useful First Test
Ask for a tiny interactive artifact that has an unambiguous check: “Build a single-page mechanical clock with labeled parts; make the hands move at the right rates, let me pause and resume, and include a short test checklist.” Save the prompt, model and effort setting. Then check the clock after a simulated hour, inspect any claimed mechanics, test keyboard use, record latency and cost, and compare one alternative model using the same brief. The model’s best-looking screenshot is not a substitute for a working result.
For vision, count only directly visible objects first. Label occluded-object estimates explicitly as estimates. For code, run the tests yourself and inspect generated dependencies before installing them. This is how to turn an exciting demo into evidence about your own workflow.
Video Chapters
3D and visual builds
| Time | Topic |
|---|---|
| 00:00 | The new model and mechanical watch |
| 02:27 | Fall Guys-style game |
| 03:43 | Foldable phone model |
| 05:59 | Multimodal pastry count |
Speed, agents and local software
| Time | Topic |
|---|---|
| 09:14 | Text-generation speed |
| 10:41 | Finding interview clips |
| 12:50 | Locally built spreadsheet |
| 14:09 | Rubik’s cube test |
| 15:06 | Personal PDF reader |
Benchmarks and price
Sources and Limits
- Video: Matthew Berman’s full test and the user-supplied transcript. Visual judgments, local hardware and third-party price/speed comparisons are his observations.
- Official release and model: DeepSeek’s 17 October announcement, model card, files, and reference inference guide. Benchmarks are vendor-reported.
- Access and setup: first API call, live pricing, Codex setup, and Hugging Face CLI.
Checked 28 October 2026. Availability, model routing, prices and runtime support can change. This article does not claim that we installed the 510 GB checkpoint or reproduced the on-screen builds.