Video summary

I Ran the FULL DeepSeek V4.1 (1017) via SSD for 24 Hrs... New #1 Local AI 🤯

Main summary

Key takeaways

Technology

Summary

The video reviews a local run of DeepSeek V4.1-1017, focusing on whether a very large model can run on an Apple Silicon workstation by offloading part of it to SSD.

Local setup, memory, and speed

  • The model is described as roughly 510 GB, with about 40% made up of “engrams” that can be offloaded to storage.
  • The host runs it on an M3 Ultra Mac Studio with 512 GB of unified memory. During live tests, DeepSeek uses around 278–280 GiB, while total system memory use is about 326 GB, including cached files.
  • Offloaded engrams are read from storage, though the operating system may cache them in RAM. This helps keep inference fast once the data is cached.
  • Generation is typically around 15 tokens per second in the demonstrated setup, including a reported 15.5 tokens/s in an early test. The host says speculative decoding/MTP was not yet enabled.
  • The host tests a Q9 version that quantizes only the vision layers, describing it as nearly lossless, and compares it with the pre-quantized version and other quantization variants.

Benchmarks and model comparisons

  • The presenter reports strong benchmark results, including a 77 score on a SWE-related benchmark, while noting that the benchmark’s special evaluation setup may produce lower results on conventional systems.
  • Other cited results include TerminalBench around 30–40 and ProgramBench around 29. The host compares these with lower scores attributed to Kimi K3 and GLM 5.3.
  • These are the presenter’s reported comparisons; the video does not establish that all models were tested under identical conditions.
  • A recurring observation is that quantization does not always reduce quality: in some of the host’s tests, a Q9 variant appears to perform better than the pre-quantized model.

Temperature, reasoning, and coding reliability

The host compares thinking off, low or maximum thinking, and different temperature settings.

  • Temperature 0 is presented as more reliable for code generation. It often produces working results and tends to use the model’s highest-probability token choices.
  • Temperature 1 sometimes produces more creative or impressive results, but can also cause syntax or runtime errors, choose incorrect variable names, or end its reasoning early despite a larger thinking budget.
  • The presenter notes that problems can often be fixed by supplying an error or screenshot and asking for another pass. They also discuss using separate files or tool calls for larger coding tasks rather than putting everything into one response.
  • The host experiments with reusing older reasoning traces and proposes a “continue thinking” control to extend the model’s reasoning when it stops prematurely.

Demonstrations and review tests

  • Interactive visual simulations: DeepSeek creates a nuclear-explosion scene and a volcanic eruption. The host iteratively improves the volcano by describing problems and supplying screenshots, prompting changes to lighting, shaders, and glow effects. The local volcano generation reportedly takes about 20 minutes, compared with roughly 2–3 minutes for the cloud version.
  • Image understanding: The model describes a cat and uses screenshots to diagnose and revise generated projects, including identifying a lighting and bloom issue.
  • Human visuals: It generates a face and a 3D anatomy visualization. The presenter considers the anatomy result particularly strong, with recognizable organs and a connected skeleton.
  • Web design: Initial website results are judged plain; a more elaborate prompt produces a more polished, interactive interface. One higher-temperature attempt fails with a runtime error.
  • Games and interactive software: The host tests GTA-style driving, Age of Empires-like scenes, Gears of War, Doom, Street Fighter, Tomb Raider, and an Assassin’s Creed-style parkour game. Results vary widely: some are incomplete or buggy, while the Tomb Raider and parkour demonstrations are praised for their movement and game mechanics. Doom improves after the host points out that the first version lacks a weapon.
  • Creative applications: DeepSeek builds a simple Blender-like 3D editor with primitives and object manipulation, an iMovie-like video editor interface, and a Logic Pro-style music tool that can generate audio. The presenter describes these as promising prototypes rather than full replacements for the original software.
  • Logic and programming: In a WinRT subscription question, the model gives the correct answer that the relevant product type may be reported as “durable,” including with a limited thinking budget. The host contrasts this with another model’s incorrect low-reasoning answer.

Overall assessment

The presenter’s main conclusion is that SSD-backed engram offloading makes a very large model usable locally, with substantial memory savings and reasonable generation speed. DeepSeek V4.1-1017 shows strong coding and multimodal prototyping potential, but its output quality is inconsistent. Careful prompting, iterative fixes, temperature choice, and quantization can substantially affect results.

Local execution is slower than the cloud for some tasks, but the host highlights privacy and infrastructure control as potential benefits.

Main speaker/source: A single host from the xCreate channel, presenting the review and running the local-model demonstrations.

Rate this summary

Your feedback will help improve summaries.

Improve this summary

Reprocess with a stronger model when the summary feels incomplete or inaccurate.

Pro

Translate summary in another language

Pro

Ask questions to this video

Chat for follow-up questions, clarifications, and source-backed answers.

Coming soon

Share this summary

Original video