DeepSeek R1 7B vs 8B: Which Distilled Model Should You Run?
On this page
If you open Ollama or LM Studio and search for DeepSeek R1, you will see a 7B and an 8B version sitting next to each other. They look like two sizes of the same model, but they are not. Both are small "distilled" models built on top of other companies' open models, and the 8B tag in Ollama has even changed what it points to since early 2025.
This guide explains what each one really is, how they score, and which one to download.
Quick answer
- Using Ollama? Pick
deepseek-r1:8b. It now downloads DeepSeek-R1-0528-Qwen3-8B, a newer distill that scores far higher than the original 7B and 8B on DeepSeek's published benchmarks. - Mostly maths problems on a tight GPU? The 7B (DeepSeek-R1-Distill-Qwen-7B) is slightly smaller and was built on a maths-focused base model.
- Need the original Llama-based 8B (for example because your tools expect Llama 3.1)? Download DeepSeek-R1-Distill-Llama-8B from Hugging Face and keep in mind it follows the Llama 3.1 licence.
Neither one is the "real" DeepSeek R1
The full DeepSeek R1 is a 671-billion-parameter mixture-of-experts model that needs a server full of GPUs (the full model is a 404 GB download in Ollama). To make R1's reasoning usable on normal PCs, DeepSeek generated around 800,000 training examples with R1 and used them to fine-tune smaller open models. That process is called distillation, and the results are the 1.5B, 7B, 8B, 14B, 32B and 70B "R1" models.
So the difference between 7B and 8B is not one extra billion parameters on the same design. They are built on different base models from different companies.
What each model is built on
| R1 7B | R1 8B (original, Jan 2025) | R1 8B (0528, current Ollama tag) | |
|---|---|---|---|
| Full name | DeepSeek-R1-Distill-Qwen-7B | DeepSeek-R1-Distill-Llama-8B | DeepSeek-R1-0528-Qwen3-8B |
| Base model | Qwen2.5-Math-7B (Alibaba) | Llama 3.1 8B (Meta) | Qwen3 8B Base (Alibaba) |
| Parameters | 7.6B | 8.0B | 8B |
| Licence | MIT (base model Apache 2.0) | MIT, plus Llama 3.1 licence terms | MIT |
| Ollama tag | deepseek-r1:7b (4.7 GB) | No longer the default 8b tag | deepseek-r1:8b (5.2 GB) |
Benchmark results
These scores come from DeepSeek's own model cards on Hugging Face (pass@1, higher is better). The original 7B and 8B are almost tied: the 7B is better at maths, the 8B slightly better at coding.
| Benchmark | R1 Distill Qwen 7B | R1 Distill Llama 8B |
|---|---|---|
| AIME 2024 (competition maths) | 55.5 | 50.4 |
| MATH-500 | 92.8 | 89.1 |
| GPQA Diamond (science questions) | 49.1 | 49.0 |
| LiveCodeBench (coding) | 37.6 | 39.6 |
| Codeforces rating | 1189 | 1205 |
The newer DeepSeek-R1-0528-Qwen3-8B is in a different league. DeepSeek reports 86.0 on AIME 2024, 61.1 on GPQA Diamond and 60.5 on LiveCodeBench, which is better than the plain Qwen3-8B on AIME (76.0) and far ahead of both original distills. Treat all of these as the publisher's own numbers; small models vary a lot from one prompt to the next, so test with your own tasks.
Hardware, speed and memory
Because the two models are almost the same size, the hardware you need is nearly identical. The default 4-bit downloads in Ollama are 4.7 GB (7B) and 5.2 GB (8B).
- GPU with 8 GB of VRAM or more: either model fits fully on the GPU and runs smoothly.
- GPU with 6 GB: both will run, but part of the model may spill to system RAM and slow down, more so with long conversations.
- No GPU: both run on the CPU with 16 GB of RAM, just a lot more slowly.
Speed differences between them are small; the 7B is marginally faster simply because it is a little smaller. Remember that R1 models "think" before answering, so even a fast model can take a while on hard questions. Our DeepSeek R1 hardware requirements guide covers the bigger versions, and LM Studio vs Ollama helps you choose an app to run them in.
Can you use them through the DeepSeek API?
No. DeepSeek's own API and chat app run the full-size models, not the 7B or 8B distills. The distills are meant to be downloaded and run locally (or on a host that offers open-weight models), so there is no DeepSeek price difference between them.
Settings that make a difference
DeepSeek's model cards recommend a few settings for these distilled models:
- Set the temperature between 0.5 and 0.7 (0.6 is the suggested value) to avoid endless repetition.
- Don't use a system prompt; put all instructions in your normal message.
- For maths, ask the model to reason step by step and put the final answer in
\boxed{}.
Which one should you choose?
- Most people:
ollama run deepseek-r1:8b. You get the newer Qwen3-based distill, which beats both originals on DeepSeek's maths, science and coding benchmarks for only 0.5 GB more. - Older guides or scripts that say "7B": the 7B is still a reasonable choice for maths-heavy work and slightly smaller downloads.
- Want stronger reasoning and have 12 GB or more of VRAM? Step up to the 14B; the jump in quality is bigger than the jump from 7B to 8B.
FAQ
Is DeepSeek R1 8B better than 7B?
The current Ollama 8B (DeepSeek-R1-0528-Qwen3-8B) is clearly better on DeepSeek's published benchmarks. The original January 2025 Llama-based 8B was roughly equal to the 7B: weaker at maths, slightly stronger at coding.
Why does my 8B download say Qwen3 instead of Llama?
Ollama switched its deepseek-r1:8b tag to the DeepSeek-R1-0528-Qwen3-8B model after DeepSeek released it. If you need the Llama version specifically, download DeepSeek-R1-Distill-Llama-8B from Hugging Face.
Can I use these models commercially?
DeepSeek releases the weights under the MIT licence. The original 8B is also bound by Meta's Llama 3.1 licence, so check its terms before using it in a product.