DeepSeek R1 Tested on Coding, Reasoning and Sensitive Questions

I am testing the full DeepSeek R1 reasoning model against my usual collection of prompts. Vultr sponsors this video and provides the cloud GPUs used for the demonstration. The full model has 671 billion total parameters; DeepSeek reports roughly 37 billion active per token in its mixture-of-experts architecture.

The model runs on a cloud machine through Open WebUI, rather than DeepSeek’s own hosted chat interface. As a basic check, I ask how many times the letter R appears in “strawberry.” The model displays text inside thinking tags and correctly answers three.

I find the visible back-and-forth interesting because the generated text revisits and checks its answer. It reads to me like a human internal monologue, although the text alone does not establish that its internal process is human-like or that every explanation faithfully represents how the answer was produced.

For a first coding challenge, I ask for Snake in Python. R1 outlines an approach using Pygame, then outputs code and instructions. I copy the program into my local editor and run it: the movement, score and restart behavior appear to work in this demonstration, so I give it a pass.

Next I ask for Tetris, a harder task in my personal rubric. The model spends time laying out pieces, collisions and rotation before producing a 179-line program. I am impressed by the attention to edge cases, although this is a single output rather than a controlled comparison of models.

After several minutes of generation, I run the program. Pieces appear and completed lines clear; I notice it does not keep score or preview the next piece. I still count the example as a pass because the main game works in my brief test.

The sponsored Vultr system shown has 128 CPU cores and eight AMD Instinct GPUs, each with 192 GB of GPU memory. That describes the hardware used in this run, not the only possible configuration for R1 or smaller distilled models. The promotion mentioned in the video was a January 2025 offer.

For a postal-envelope puzzle, the model converts a 200-by-275-millimeter envelope to 20 by 27.5 centimeters and considers how to orient it against the stated minimum and maximum dimensions. It concludes that the envelope fits, which I count as a pass for the dimensions given in the prompt.

I ask the self-referential question of how many words are in the response. The model considers several possible wordings and answers with a seven-word sentence, followed by an explanation. I count that as a pass by treating the sentence as the answer, though counting the entire response could lead to a different judgment.

In the three-killers puzzle, the model reasons that the person entering and killing one of the original three becomes a killer too. It answers three living killers, while discussing possible ambiguity about counting the deceased person. I like that the displayed reasoning considers the assumptions rather than only producing a number.

The model also answers the inverted glass and marble puzzle: when the glass is lifted and moved, the marble remains on the table under the ordinary assumptions. In a decimal comparison, it correctly identifies 9.9 as larger than 9.11 by writing 9.9 as 9.90.

I then test politically sensitive questions in this self-hosted deployment. It refuses a question about Tiananmen Square and gives a one-sided response reflecting the Chinese government’s position when asked about Taiwan. These are outputs for particular prompts, not a complete assessment of the model’s political behavior.

I had heard that restrictions might apply only to the hosted service, but this deployment also produces a refusal. I speculate that the Taiwan response may be hardcoded because the displayed thinking section is empty. The demonstration does not establish whether the behavior comes from model training, prompts, serving software or another part of the configuration.

I contrast this with a separate request about committing a crime and observe that the model appears willing to answer. A refusal to provide harmful instructions is a safety behavior, not automatically a reasoning failure; one prompt also cannot establish the model’s overall safety profile.

For the final prompt, I ask for ten sentences ending in “apple,” and the displayed results satisfy that requirement. I come away impressed by these selected tests and the visible reasoning. My description of the performance as flawless is a personal assessment of a small sample, not a comprehensive measurement.