[whatmodel.sh](../../README.md) / DeepSeek R1 Distill Llama 70B
# DeepSeek R1 Distill Llama 70B
DeepSeek · released 2025-01 · 70B params · MIT; underlying Llama 3.3 licence applies to weights
**score 58.0** ±0.9 · #225 of 732 · from 18 benchmark results across 7 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 70B among every model on that task or test, worst to best.
task vs the others rank counts tests
coding ▁▁▁▃▅▇●▇▅▄▃▁ #252 of 469 15% 5
math ▁▁▃▅█▆●█▄▂▁▁ #169 of 431 36% 5
coding · #252 of 469 · 15% of its score
Beats [Qwen3-VL 32B Instruct](qwen3-vl-32b-instruct.md), [Solar Open 100B (102B-A12B)](solar-open-100b.md), [GLM-4.7-Flash](glm-4-7-flash.md) here, though they rank higher overall. Trails [QwQ-32B-Preview](qwq-32b-preview.md), [InternLM2.5 20B Chat](internlm2-5-20b-chat.md), [Ministral 3 3B Reasoning 2512](ministral-3-3b-reasoning.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
LiveCodeBench (version unspecifie… 57.5% ▄····▄▄··▄●▄ beats 5/6 medium · avg 42% 6.9% at its level
CodeReviewQA: change localisation… 47.3% █▂▁▁▂▂▂▃▅▅●▇ beats 35/47 hard · avg 33% 3.4% sharp · at its level · noisy
CodeReviewQA: automated code refi… 28.7% ▁▁▁▁▁▁▂▂▃●▅▄ beats 19/47 hard · avg 32% 2.7% at its level
CodeReviewQA: solution identifica… 68.8% ▃··▁▂▂▃▂█●▄▂ beats 37/47 medium · avg 51% 2.1% sharp · noisy
CodeReviewQA: change type recogni… 61.2% ▁▁▁······▁●█ beats 11/47 easy · avg 67% 0.1% blunt · at its level · noisy
math · #169 of 431 · 36% of its score
Beats [MiMo-V2.5 (310B-A15B)](mimo-v2-5.md), [MiniMax-M2.7](minimax-m2-7.md), [MiniMax-M2.1](minimax-m2-1.md) here, though they rank higher overall. Trails [Olmo 3 7B Think](olmo-3-7b-think.md), [Nemotron Nano 9B v2](nemotron-nano-9b-v2.md), [Qwen3-VL 4B Thinking](qwen3-vl-4b-thinking.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
MATH level 5 89.9% ▂▂▅▆▄█▆▃▅●▃▄ beats 78/92 hard · avg 28% 18% sharp · at its level
OTIS mock AIME 2024–2025 51.4% ▃▃▄▃▃▃●▇█▃▂▅ beats 78/197 very hard · avg 7.2% 6.7% sharp · at its level · noisy
AIME 2024 (maker) 70.0% ▃·▃·▃▃···●·▃ beats 5/7 medium · avg 43% 5.2% sharp · at its level · noisy · several …
AIME 2025 51.5% ●··▄··█··▄·▄ beats 0/6 hard · avg 32% 4.6% sharp · at its level · noisy
MATH-500 (maker) 94.5% ▄·▄▄·▄▄●···▄ beats 6/7 very easy · avg 91% 1.6% maxed out
other tests (5)
test result vs the others in its run difficulty counts why
GPQA Diamond 55.7% ▁▁▁▁▂▄▆●▅▇█▅ beats 79/211 hard · avg 39% 38% sharp · at its level · several runs
BRIDGE clinical text, zero-shot a… 39.8% ▁▁▁▁▁▂▂▃▄▆●▃ beats 70/97 hard · avg 31% 5.7% blunt · at its level
Judgemark v2 (LLM-as-judge for cr… 59.7% ▃▃··▃●▅···█▃ beats 3/10 medium · avg 53% 2.9% at its level
EuroEval HellaSwag (English), tes… 39.8% ▃·▁▁▂●▂▅█▇▄▁ beats 21/100 medium · avg 67% 1.3% sharp · noisy
Omni-MATH 63.4% █·●·█·██··██ beats 1/6 medium · avg 64% 0.5% blunt · at its level · noisy
source results share of the score
Epoch AI 3 40% ▇▇▇▇▇
Microsoft 4 23% ▇▇▇
deepseek-ai/DeepSeek-R1 model card (sub… 4 18% ▇▇
CodeReviewQA: The Code Review Comprehen… 4 8% ▇
BRIDGE Medical Leaderboard 1 6% ▇
Judgemark v2 results 1 3% ▇
EuroEval leaderboards 1 1% ▇
prior: pull toward the middle 0% ▇
weighted average = 58.0
Licences: Epoch AI ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · Microsoft ([CC BY 4.0](../../licenses/report-2504.21318v1.txt)) · deepseek-ai/DeepSeek-R1 model card ([mit](../../licenses/deepseek-ai--DeepSeek-R1.txt)) · CodeReviewQA: The Code Review Comprehension Assessment for Large Language Models ([CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/)) · BRIDGE Medical Leaderboard ([MIT](../../licenses/bridge-medical-leaderboard-MIT.txt)) · Judgemark v2 results ([MIT](../../licenses/judgemark-v2-MIT.txt)) · EuroEval leaderboards ([MIT](https://github.com/EuroEval/leaderboards/blob/1a77f1d2f8ba1ba05ab73287fb090d76bf41b2e0/LICENSE))
## How to run it
### Through an API
api (Novita AI) · $0.80/M blended · $0.8 in / $0.8 out per 1M tokens · [novita.ai](https://novita.ai/models/llm/deepseek-deepseek-r1-distill-llama-70b)
[price source](https://novita.ai/pricing): USD per 1M tokens, Novita AI serverless chat completions, model deepseek/deepseek-r1-distill-llama-70b (status active): input $0.80, output $0.80. Novita serves it with an 8,192-token context (its own serving limit). Currency: USD. Novita AI official pricing page, observed 2026-09-25. · [context source](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B/blob/b1c0b44b4369b597ad119a196caf79a9c40e141e/config.json): DeepSeek official configuration for DeepSeek-R1-Distill-Llama-70B (pinned commit): max\_position\_embeddings 131072 (llama3 rope scaling, factor 8 over original 8192). Configured maximum, not a benchmarked long-context claim: the official card states no context length for the distilled models (its table gives 128K only for DeepSeek-R1/R1-Zero), reports a 32,768-token maximum generation length, and its vLLM example uses --max-model-len 32768. Observed 2026-09-25.
### 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 70B (Q8_0, needs ~74.98 GB, from [huggingface.co/unsloth](https://huggingface.co/unsloth/DeepSeek-R1-Distill-Llama-70B-GGUF/tree/732dd974083ea5877d7b6d788b36fe7c2e5eab36))
```sh
llama cli -hf unsloth/DeepSeek-R1-Distill-Llama-70B-GGUF:Q8_0
```
3. Or start a local server and open the address it prints (http://localhost:9931)
```sh
llama serve -hf unsloth/DeepSeek-R1-Distill-Llama-70B-GGUF:Q8_0
```
Other versions: ud-q2_k_xl (~26.97 GB) `llama cli -hf unsloth/DeepSeek-R1-Distill-Llama-70B-GGUF:UD-Q2_K_XL`
### 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 70B (4bit, needs ~39.7 GB, Mac only, from [huggingface.co/mlx-community](https://huggingface.co/mlx-community/DeepSeek-R1-Distill-Llama-70B-4bit/tree/6b0c127e00720c480ae36462bb15e71eeb521e80))
```sh
lms get https://huggingface.co/mlx-community/DeepSeek-R1-Distill-Llama-70B-4bit -y
```
3. Then run lms ls to see its name, and lms chat followed by that name
Other versions: Q8_0 (~74.98 GB) `lms get https://huggingface.co/unsloth/DeepSeek-R1-Distill-Llama-70B-GGUF@Q8_0 -y` · ud-q2_k_xl (~26.97 GB) `lms get https://huggingface.co/unsloth/DeepSeek-R1-Distill-Llama-70B-GGUF@UD-Q2_K_XL -y` · 8bit (~75 GB, Mac only) `lms get https://huggingface.co/mlx-community/DeepSeek-R1-Distill-Llama-70B-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 70B (Q4_K_M, needs ~42.52 GB, from [ollama.com](https://ollama.com/library/deepseek-r1/tags))
```sh
ollama run deepseek-r1:70b
```
### 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 70B (4bit, needs ~39.7 GB, Mac only, from [huggingface.co/mlx-community](https://huggingface.co/mlx-community/DeepSeek-R1-Distill-Llama-70B-4bit/tree/6b0c127e00720c480ae36462bb15e71eeb521e80))
```sh
mlx_lm.chat --model mlx-community/DeepSeek-R1-Distill-Llama-70B-4bit
```
3. Or start an OpenAI-compatible server at http://localhost:8080/v1
```sh
mlx_lm.server --model mlx-community/DeepSeek-R1-Distill-Llama-70B-4bit
```
Other versions: 8bit (~75 GB, Mac only) `mlx_lm.chat --model mlx-community/DeepSeek-R1-Distill-Llama-70B-8bit`
Model repository and documentation: [huggingface.co/deepseek-ai](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B)
## 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) Average of AIME 2024, FOLIO and T… 0.818 → 0.804 98.3%
AWQ 4-bit (AutoAWQ; vLLM) AIME 2024 accuracy (mean of 3 pas… 0.656 → 0.634 96.6%
FP8 W8A8 dynamic (FP8 weights… AIME 2024 pass@1 0.678 → 0.692 102.0%
FP8 W8A8 dynamic (FP8 weights… Reasoning average of 3 tasks (AIM… 0.762 → 0.765 100.3%
FP8 W8A8 dynamic (FP8 weights… GSM8K 5-shot strict-match (lm-eva… 0.93 → 0.93 100.0%
FP8 W8A8 dynamic (FP8 weights… MATH-500 pass@1 0.953 → 0.951 99.8%
FP8 W8A8 dynamic (FP8 weights… OpenLLM v1 average of 6 tasks (AR… 0.757 → 0.755 99.7%
FP8 W8A8 dynamic (FP8 weights… GPQA Diamond pass@1 0.656 → 0.651 99.4%
FP8 W8A8 dynamic (FP8 weights… MMLU 5-shot acc (lm-evaluation-ha… 0.78 → 0.774 99.2%
FP8 W8A8 dynamic (FP8 weights… IFEval 0-shot instruction-level s… 0.424 → 0.421 99.2%
FP8 W8A8 dynamic (FP8 weights… OpenLLM v2 average of 6 tasks (IF… 0.154 → 0.149 96.6%
GPTQ 3-bit (vLLM) Average of AIME 2024, FOLIO and T… 0.818 → 0.726 88.8%
GPTQ 3-bit (vLLM) AIME 2024 accuracy (mean of 3 pas… 0.656 → 0.467 71.2%
GPTQ 4-bit (vLLM) AIME 2024 accuracy (mean of 3 pas… 0.656 → 0.667 101.7%
GPTQ 4-bit (vLLM) Average of AIME 2024, FOLIO and T… 0.818 → 0.812 99.3%
INT4 W4A16 (weights only, sym… OpenLLM v1 average of 6 tasks (AR… 0.757 → 0.759 100.2%
INT4 W4A16 (weights only, sym… GSM8K 5-shot strict-match (lm-eva… 0.93 → 0.93 100.0%
INT4 W4A16 (weights only, sym… IFEval 0-shot instruction-level s… 0.424 → 0.424 99.9%
INT4 W4A16 (weights only, sym… MATH-500 pass@1 0.953 → 0.952 99.9%
INT4 W4A16 (weights only, sym… MMLU 5-shot acc (lm-evaluation-ha… 0.78 → 0.771 98.9%
INT4 W4A16 (weights only, sym… Reasoning average of 3 tasks (AIM… 0.762 → 0.75 98.3%
INT4 W4A16 (weights only, sym… GPQA Diamond pass@1 0.656 → 0.64 97.7%
INT4 W4A16 (weights only, sym… AIME 2024 pass@1 0.678 → 0.656 96.7%
INT4 W4A16 (weights only, sym… OpenLLM v2 average of 6 tasks (IF… 0.154 → 0.136 88.1%
INT8 W8A8 (weights symmetric … GSM8K 5-shot strict-match (lm-eva… 0.93 → 0.936 100.6%
INT8 W8A8 (weights symmetric … OpenLLM v1 average of 6 tasks (AR… 0.757 → 0.76 100.4%
INT8 W8A8 (weights symmetric … IFEval 0-shot instruction-level s… 0.424 → 0.425 100.1%
INT8 W8A8 (weights symmetric … MATH-500 pass@1 0.953 → 0.953 100.0%
INT8 W8A8 (weights symmetric … MMLU 5-shot acc (lm-evaluation-ha… 0.78 → 0.78 100.0%
INT8 W8A8 (weights symmetric … AIME 2024 pass@1 0.678 → 0.678 99.9%
INT8 W8A8 (weights symmetric … Reasoning average of 3 tasks (AIM… 0.762 → 0.76 99.7%
INT8 W8A8 (weights symmetric … GPQA Diamond pass@1 0.656 → 0.65 99.1%
INT8 W8A8 (weights symmetric … OpenLLM v2 average of 6 tasks (IF… 0.154 → 0.152 99.0%
Each kept value links to its source.
[interactive version](https://whatmodel.sh/model/deepseek-r1-distill-llama-70b) · [JSON](https://whatmodel.sh/api?format=json&all=1)