DeepSeek-R1: the repository is a README and a PDF, and the headline claim belongs to another model

The repository is a README, a PDF and a figures folder

Five entries sit at the top level: .github/, DeepSeek_R1.pdf, LICENSE, README.md and a figures directory. There is no inference script, no requirements file, no tokenizer, no config directory and no example. The weights are not here either, because the download table sends every reader to a HuggingFace repository instead.

Consequence: cloning this gives you a document, not a program. Anyone expecting a git clone to produce a working local model ends up with the runnable artefact split across two places, checkpoints on HuggingFace and architecture in a separate repository, and neither is resolved from here. The last commit to main is dated 2025-06-27 and the project has exactly one release, v1.0.0, from the same day.

R1-Zero is the experiment, and its three named failures are why R1 exists

Two models are published. DeepSeek-R1-Zero was trained through large-scale reinforcement learning with no supervised fine-tuning as a preliminary step, and the README credits that with reasoning behaviours that emerged on their own, among them self-verification, reflection and long chains of thought. The same passage names what went wrong: endless repetition, poor readability and language mixing.

Consequence: R1-Zero is a research artefact and the README says as much, yet the download table presents it beside R1 with identical shape, 671B total, 37B activated and 128K context, so nothing in the table separates a usable model from a documented failure. Pick the Zero checkpoint expecting R1's behaviour and you inherit the three problems its own authors listed, and the fix for them is the other download.

The pure-RL result belongs to R1-Zero, and R1 got supervised data first

The pipeline section splits two stories apart. R1-Zero applies reinforcement learning straight to the base model, and the README calls it the first open research validating that an LLM's reasoning can be incentivised purely through RL with no SFT at all. R1 is described differently: two RL stages, one to discover improved reasoning patterns and one to align with human preferences, wrapped around two SFT stages that seed reasoning and non-reasoning behaviour, fed by cold-start data placed before the reinforcement learning.

Consequence: the sentence that gets quoted is about a model nobody should deploy. R1, the checkpoint with the same parameter counts, went through supervised fine-tuning, so reading trained purely by reinforcement learning as a property of DeepSeek-R1 contradicts the repository's own account. If the absence of SFT is what you care about, R1-Zero is the artefact that has it, along with the repetition and the language mixing.

671B total and 37B active, with the architecture documented elsewhere

Both full checkpoints are mixture-of-experts models with 671B total parameters, 37B activated parameters and a context length of 128K. Neither is described here. The README states that R1-Zero and R1 are trained on DeepSeek-V3-Base and sends architecture questions to the DeepSeek-V3 repository.

Consequence: you cannot size hardware or plan a serving configuration from this repository, because the parts that decide memory per token, the router, the expert layout and the parallelism strategy, all live in another project. The six distilled dense models come with plain parameter figures and no such question attached, which is much of why people end up running one of those instead.

The distills changed their configs and tokenizers, and you are told not to reuse the base model's

Six dense checkpoints are published, from 1.5B to 70B, on the bases Qwen2.5-Math-1.5B, Qwen2.5-Math-7B, Llama-3.1-8B, Qwen2.5-14B, Qwen2.5-32B and Llama-3.3-70B-Instruct. Two bases are maths-specific and the largest is an instruct-tuned Llama, so the family is not one recipe applied six times. Two sentences carry the operational warning: the distills are fine-tuned using samples generated by DeepSeek-R1, and their configs and tokenizers were changed slightly, with the instruction to use the project's setting to run them.

Consequence: loading a distill with the stock config of its base model is the default mistake, and it fails quietly, because the architecture still matches and the output simply degrades. No diff of what changed is offered, so the setting has to be taken from the checkpoint rather than reconstructed from the base.

Every score comes from 64 samples per query, not from one run

The evaluation section fixes its protocol before it prints anything. For all models the maximum generation length is set to 32,768 tokens, and for benchmarks requiring sampling the temperature is 0.6, the top-p value is 0.95, and 64 responses are generated per query to estimate pass@1.

Consequence: the headline figures are best-of-64 statistics rather than what a first attempt returns, so a single-sample run of the same checkpoint lands below them and that gap is not a defect. It also means reproducing the table costs 64 generations per prompt, and that 32,768 tokens is a hard ceiling on how long a chain of thought can run before it is cut off mid-thought.

On MMLU, R1 sits below o1-1217 while the summary says comparable

The comparison table is the only place the README makes a head-to-head claim, and its first English row is more precise than the prose around it. On MMLU pass@1, Claude-3.5-Sonnet-1022 scores 88.3, GPT-4o 0513 scores 87.2, DeepSeek V3 scores 88.5, OpenAI o1-mini scores 85.2, OpenAI o1-1217 scores 91.8 and DeepSeek R1 scores 90.8. The prose summary is that R1 achieves performance comparable to OpenAI-o1 across math, code and reasoning tasks, and separately that Distill-Qwen-32B outperforms o1-mini across various benchmarks.

Consequence: comparable is carrying weight in that sentence. On the one row visible here R1 sits a point under o1-1217 and above o1-mini, which is a defensible summary and not a win, and a reader who takes the headline instead of the row will overestimate it. Everything in the repository dates from 2025-06-27, so none of these figures describe a model you would be racing today.

Editorial conclusion

Use DeepSeek-R1 when you want reasoning weights that read and reason under their own settings, and reach for a distill when you need something one machine can hold. Two things the README states and most summaries skip. The distilled checkpoints ship modified configs and tokenizers, and the project asks you to use their setting rather than the base model's. And the evaluation numbers come from 64 sampled responses per query at temperature 0.6 under a 32,768 token generation cap, so they will not line up with a single-shot run of your own. This repository has not been touched since 2025-06-27 and carries a single v1.0.0 release, so read it as a frozen record of that model rather than as a project in motion.

Frequently asked questions

Official sources

1. Official README
2. Project repository
3. Release notes