DeepSeek-R1: open weights that reason

On 20 January 2025 DeepSeek released R1. The specialty is not a new chat skin. It is a reasoning model whose weights you can download, under an MIT licence, with scores they put next to OpenAI’s o1.

This post is what they opened, which numbers they published, and how a distilled small model changes a student project.

What launched

DeepSeek-R1 is a large mixture-of-experts model (they list 671B total parameters, about 37B active). They trained it with a lot of reinforcement learning after the base, and said they needed little labelled data for the jump on math and code.

They also released R1-Zero and six distilled models (1.5B to 70B) so you can run a smaller copy. The 32B and 70B distill, they said, sat near o1-mini on their tables.

You could try “DeepThink” on their chat site, call deepseek-reasoner on the API, or take the weights. API list price at launch was $0.14 per million input tokens on a cache hit, $0.55 on a miss, $2.19 per million output.

If you can run the small distill on your five cases, you do not need the giant for that step.

What the specialty is

Open reasoning. You can inspect, distill, and fine-tune. You are not only renting a picker.

That matters when you want a local or cheap path for contest-style questions, or you want to study how a reasoner writes its chain.

It does not mean “always better than a closed model on your product.” On SWE-bench Verified they reported about 49%, close to other models in that table, not a coding crown. The crown they claimed was math and code-contest scores.

Numbers they published

From their paper table (R1 vs the copies they listed):

The distill table is the useful one for a laptop: their Qwen-32B distill was 72.6% AIME pass@1 in that same release. Still far above GPT-4o’s 9.3% on that exam, at a size you might host.

Again: their eval, their prompts, their temperature. Copy the idea, not the trophy.

Walk through once

You want a helper that explains a worked percentage, then a helper that solves a contest-style algebra item.

For the percentage: a 7B or even no model. Code first.

For the algebra item: try their 32B distill if you can host it, or the API reasoner. Write the item and the numeric answer. Run three times. If it flakes, you have a temperature and check problem, not a “need the 671B” problem.

A bait: “Cite the hostel fee from this empty folder.” Open weights do not invent a file. Refuse.

If you fine-tune a distill on your chat logs, you can copy jokes and numbers you should not store. Distill is a starting point, not a dump.

What good looks like

You can name the date and the licence.

You have one contest case that needs a reasoner and four cases that do not.

A 32B distill, or the API, passes the contest case often enough that you wrote the flake down. You did not host 671B to feel serious.

The class list is not in the training folder.

The crate is the release. Most student projects need the small chip they poured off.

What open actually gives you

You can read a paper and a repo. You can see a chain they did not fully hide. You can distill again.

You still need eval. An open model that fails your bait is not safer because the file is on disk.

You still need a licence read. MIT is wide. Your college data may not be. Do not pour student rolls into a public fine-tune.

Common mistakes

Treating “on par with o1” as a default picker. On par on AIME is not on par on your form.

Skipping the distill. You will try to host 671B and quit.

No cache. Their cheap input number was the cache hit.

Fine-tuning on secrets. Open is not a privacy policy.

Ignoring SWE-bench. If the job is a repo edit, this launch is not the specialty.

Practise

1. Write five cases: two easy counts, one contest item, one empty-file refuse, one joke.
2. Run them on a small distill or the API reasoner.
3. Record pass/fail, not a vibe.
4. If only the contest item needs the reasoner, route that one.
5. Do not upload a class list to a public host.

Start mid. Climb only when a named case fails.

Checklist

• You can date the release: 20 January 2025.
• You can name the specialty: open reasoning plus distills.
• AIME / MATH numbers are labelled as theirs.
• A small distill is the first host you try.
• Student data stays off a public fine-tune.

The next step is five cases on a distill, not a poster about beating a lab.