[whatmodel.sh](../../README.md) / DeepSeek R1 Distill Llama 8B
# DeepSeek R1 Distill Llama 8B
DeepSeek · released 2025-01 · 8.03B params · mit
**score 46.4** ±1.2 · #446 of 732 · from 19 benchmark results across 8 sources
Score within the [LMArena (text)](https://whatmodel.sh/?set=lmarena-text&score=set) comparison set only.
## What it's good at
● marks DeepSeek R1 Distill Llama 8B among every model on that task or test, worst to best.
task vs the others rank counts tests
coding ▁▁▁●▅▇█▇▅▄▃▁ #440 of 469 23% 5
math ▁▁▃▅█●▄█▄▂▁▁ #213 of 431 33% 3
coding · #440 of 469 · 23% of its score
Beats [Qwen3.5 2B](qwen3-5-2b.md) here, though they rank higher overall. Trails [Llama 3.2 1B Instruct](llama-3-2-1b.md), [SmolLM2 1.7B Instruct](smollm2-1-7b.md), [Mistral 7B Instruct v0.1](mistral-7b-instruct-v0-1.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
LiveCodeBench (version unspecifie… 39.6% ▄····▄●··▄█▄ beats 2/6 medium · avg 42% 11% at its level
CodeReviewQA: solution identifica… 6.8% ▃··▁●▂▃▂█▃▄▂ beats 6/47 medium · avg 51% 4.4% sharp · at its level · noisy
CodeReviewQA: change localisation… 0.6% ●▂▁▁▂▂▂▃▅▅▅▇ beats 7/47 hard · avg 33% 4.1% sharp · noisy
CodeReviewQA: automated code refi… 4.9% ▁▁●▁▁▁▂▂▃█▅▄ beats 3/47 hard · avg 32% 3.4% at its level
CodeReviewQA: change type recogni… 0.4% ●▁▁······▁▃█ beats 0/47 easy · avg 67% 0.1% blunt · at its level · noisy
math · #213 of 431 · 33% of its score
Beats [MiniMax-M2.1](minimax-m2-1.md), [DeepSeek V3 (Dec '24)](deepseek-v3.md), [GPT-4o (2024-08-06)](gpt-4o-2024-08-06.md) here, though they rank higher overall. Trails [Granite 8B Code Base (4K)](granite-8b-code-base-4k.md), [Mixtral 8x7B v0.1 (base)](mixtral-8x7b-v0-1.md), [Olmo 3 7B Instruct](olmo-3-7b-instruct.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
AIME 2024 (maker) 50.4% ▃·▃·●▃···█·▃ beats 2/7 medium · avg 54% 15% sharp · at its level · noisy · several …
MATH-500 (maker) 89.1% ▄·●▄·▄▄█···▄ beats 1/7 very easy · avg 91% 10% several runs
GSM8K 87.2% ▃··●··▃▃·▃·█ beats 1/7 very easy · avg 93% 8.8% sharp · noisy
other tests (5)
test result vs the others in its run difficulty counts why
GPQA Diamond (maker) 49.0% ▄·····●·█▄▄▄ beats 1/7 hard · avg 49% 27% sharp · several runs
BRIDGE clinical text, zero-shot a… 28.5% ▁▁▁▁▁▂▂●▄▆█▃ beats 24/97 hard · avg 31% 7.8% blunt · at its level
Open PT-LLM Leaderboard average (… 42.4% ▃▁▁▁▁▁●▁▂█▅▁ beats 50/276 medium · avg 72% 3.6% at its level · noisy
EuroEval HellaSwag (English), tes… 53.7% ▃·▁▁▂▁▂●█▇▄▁ beats 28/100 medium · avg 67% 2.9% sharp · at its level · noisy
EuroEval Life in the UK (English … 78.3% ▁▁▁▁▁▁▁▂▄●▄▃ beats 61/146 easy · avg 82% 2.1% at its level · noisy
source results share of the score
deepseek-ai/DeepSeek-R1 model card (sub… 4 28% ▇▇▇
microsoft/Phi-4-mini-reasoning model ca… 3 17% ▇▇
Microsoft 3 17% ▇▇
CodeReviewQA: The Code Review Comprehen… 4 12% ▇
Unsupervised Confidence Calibration for… 1 9% ▇
BRIDGE Medical Leaderboard 1 8% ▇
EuroEval leaderboards 2 5% ▇
Open Portuguese LLM Leaderboard 1 4% ▇
prior: pull toward the middle 1% ▇
weighted average = 46.4
Licences: deepseek-ai/DeepSeek-R1 model card ([mit](../../licenses/deepseek-ai--DeepSeek-R1.txt)) · microsoft/Phi-4-mini-reasoning model card ([MIT](../../licenses/card-microsoft-phi-4-mini-reasoning.txt)) · Microsoft ([CC BY 4.0](../../licenses/report-2503.01743v2.txt)) · CodeReviewQA: The Code Review Comprehension Assessment for Large Language Models ([CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/)) · Unsupervised Confidence Calibration for Reasoning LLMs from a Single Generation ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · BRIDGE Medical Leaderboard ([MIT](../../licenses/bridge-medical-leaderboard-MIT.txt)) · EuroEval leaderboards ([MIT](https://github.com/EuroEval/leaderboards/blob/1a77f1d2f8ba1ba05ab73287fb090d76bf41b2e0/LICENSE)) · Open Portuguese LLM Leaderboard ([Apache-2.0](../../licenses/hf-leaderboard-spaces-Apache-2.0.txt))
## How to run it
### llama.cpp (recommended)
1. Install llama.cpp. Homebrew and winget installs name the commands llama-cli and llama-server instead. [Other ways to install](https://github.com/ggml-org/llama.cpp/blob/master/docs/install.md)
```sh
# macOS / Linux
curl -LsSf https://llama.app/install.sh | sh
# Windows
irm https://llama.app/install.ps1 | iex
```
2. Run DeepSeek R1 Distill Llama 8B (Q4_K_M, needs ~4.9 GB, from [huggingface.co/lmstudio-community](https://huggingface.co/lmstudio-community/DeepSeek-R1-Distill-Llama-8B-GGUF/tree/d0ac80a21158e6315dccfd568bebfa5b3ed3f1d3))
```sh
llama cli -hf lmstudio-community/DeepSeek-R1-Distill-Llama-8B-GGUF:Q4_K_M
```
3. Or start a local server and open the address it prints (http://localhost:9931)
```sh
llama serve -hf lmstudio-community/DeepSeek-R1-Distill-Llama-8B-GGUF:Q4_K_M
```
Other versions: Q8_0 (~8.5 GB) `llama cli -hf lmstudio-community/DeepSeek-R1-Distill-Llama-8B-GGUF:Q8_0`
### LM Studio
1. Install LM Studio. Open the app once before using the lms command. [Download it (macOS / Windows / Linux)](https://lmstudio.ai/download)
2. Run DeepSeek R1 Distill Llama 8B (Q4_K_M, needs ~4.9 GB, from [huggingface.co/lmstudio-community](https://huggingface.co/lmstudio-community/DeepSeek-R1-Distill-Llama-8B-GGUF/tree/d0ac80a21158e6315dccfd568bebfa5b3ed3f1d3))
```sh
lms get https://huggingface.co/lmstudio-community/DeepSeek-R1-Distill-Llama-8B-GGUF@Q4_K_M -y
```
3. Then run lms ls to see its name, and lms chat followed by that name
Other versions: Q8_0 (~8.5 GB) `lms get https://huggingface.co/lmstudio-community/DeepSeek-R1-Distill-Llama-8B-GGUF@Q8_0 -y` · 4bit (~4.5 GB, Mac only) `lms get https://huggingface.co/mlx-community/DeepSeek-R1-Distill-Llama-8B-4bit -y` · 8bit (~8.5 GB, Mac only) `lms get https://huggingface.co/mlx-community/DeepSeek-R1-Distill-Llama-8B-8bit -y`
### Ollama
1. Install Ollama. [Download it (macOS / Windows)](https://ollama.com/download)
```sh
# Linux
curl -fsSL https://ollama.com/install.sh | sh
```
2. Run DeepSeek R1 Distill Llama 8B (Q4_K_M, needs ~5.32 GB, from [ollama.com](https://ollama.com/library/deepseek-r1/tags))
```sh
ollama run deepseek-r1:8b-llama-distill-q4_K_M
```
Other versions: Q8_0 (~10.14 GB) `ollama run deepseek-r1:8b-llama-distill-q8_0`
### MLX (Mac)
1. Install MLX (Mac). Apple silicon Macs only, text only.
```sh
# with uv
uv tool install mlx-lm
# or
pip install mlx-lm
```
2. Run DeepSeek R1 Distill Llama 8B (4bit, needs ~4.5 GB, Mac only, from [huggingface.co/mlx-community](https://huggingface.co/mlx-community/DeepSeek-R1-Distill-Llama-8B-4bit/tree/e23907a0aad41aa45845a938ef0fe7dfeaab3c05))
```sh
mlx_lm.chat --model mlx-community/DeepSeek-R1-Distill-Llama-8B-4bit
```
3. Or start an OpenAI-compatible server at http://localhost:8080/v1
```sh
mlx_lm.server --model mlx-community/DeepSeek-R1-Distill-Llama-8B-4bit
```
Other versions: 8bit (~8.5 GB, Mac only) `mlx_lm.chat --model mlx-community/DeepSeek-R1-Distill-Llama-8B-8bit`
Model repository and documentation: [huggingface.co/deepseek-ai](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B)
## Quantization evidence
How much of the full-precision result a quantized version kept, measured within one study. Rows from the same study share a table and aren't independent. They don't describe every download of that precision.
precision test full → quantized kept
AWQ 4-bit (AutoAWQ; vLLM) AIME 2024 accuracy (mean of 3 pas… 0.422 → 0.478 113.3%
AWQ 4-bit (AutoAWQ; vLLM) Average of AIME 2024, FOLIO and T… 0.652 → 0.666 102.1%
FP8 W8A8 dynamic (FP8 weights… AIME 2024 pass@1 0.492 → 0.508 103.2%
FP8 W8A8 dynamic (FP8 weights… IFEval 0-shot instruction-level s… 0.384 → 0.387 100.8%
FP8 W8A8 dynamic (FP8 weights… Reasoning average of 3 tasks (AIM… 0.629 → 0.633 100.6%
FP8 W8A8 dynamic (FP8 weights… MATH-500 pass@1 0.902 → 0.902 100.1%
FP8 W8A8 dynamic (FP8 weights… GPQA Diamond pass@1 0.493 → 0.487 98.9%
GPTQ 3-bit (vLLM) Average of AIME 2024, FOLIO and T… 0.652 → 0.478 73.3%
GPTQ 3-bit (vLLM) AIME 2024 accuracy (mean of 3 pas… 0.422 → 0.111 26.3%
GPTQ 4-bit (vLLM) AIME 2024 accuracy (mean of 3 pas… 0.422 → 0.422 100.0%
GPTQ 4-bit (vLLM) Average of AIME 2024, FOLIO and T… 0.652 → 0.581 89.1%
INT4 W4A16 (weights only, sym… GSM8K 5-shot strict-match (lm-eva… 0.628 → 0.631 100.5%
INT4 W4A16 (weights only, sym… MATH-500 pass@1 0.902 → 0.899 99.7%
INT4 W4A16 (weights only, sym… IFEval 0-shot instruction-level s… 0.384 → 0.379 98.6%
INT4 W4A16 (weights only, sym… OpenLLM v1 average of 6 tasks (AR… 0.599 → 0.583 97.3%
INT4 W4A16 (weights only, sym… Reasoning average of 3 tasks (AIM… 0.629 → 0.611 97.2%
INT4 W4A16 (weights only, sym… GPQA Diamond pass@1 0.493 → 0.471 95.6%
INT4 W4A16 (weights only, sym… AIME 2024 pass@1 0.492 → 0.463 94.0%
INT4 W4A16 (weights only, sym… MMLU 5-shot acc (lm-evaluation-ha… 0.556 → 0.521 93.6%
INT8 W8A8 (weights symmetric … IFEval 0-shot instruction-level s… 0.384 → 0.384 100.1%
INT8 W8A8 (weights symmetric … MATH-500 pass@1 0.902 → 0.9 99.8%
INT8 W8A8 (weights symmetric … MMLU 5-shot acc (lm-evaluation-ha… 0.556 → 0.555 99.8%
INT8 W8A8 (weights symmetric … AIME 2024 pass@1 0.492 → 0.491 99.6%
INT8 W8A8 (weights symmetric … Reasoning average of 3 tasks (AIM… 0.629 → 0.627 99.6%
INT8 W8A8 (weights symmetric … OpenLLM v1 average of 6 tasks (AR… 0.599 → 0.595 99.3%
INT8 W8A8 (weights symmetric … GSM8K 5-shot strict-match (lm-eva… 0.628 → 0.621 98.9%
Each kept value links to its source.
[interactive version](https://whatmodel.sh/model/deepseek-r1-distill-llama-8b) · [JSON](https://whatmodel.sh/api?format=json&all=1)