whatmodel.sh / DeepSeek R1 Distill Llama 70B

DeepSeek R1 Distill Llama 70B

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

DeepSeek R1 Distill Llama 70B is an open-weights language model from DeepSeek, released January 2025. Its published size is 70B parameters and a 131k-token context window. Licence: MIT; underlying Llama 3.3 licence applies to weights.

Model information sources

· Parameter count ↗

Maker states '70B' (based on Llama-3.3-70B-Instruct). The official repo's safetensors header total (HF API safetensors.total, sha b1c0b44) is 70,553,706,496 = 70.55B; sibling catalogue entries (e.g. deepseek-r1-distill-llama-8b = 8.03) use the exact header totals, so 70.55 may be preferred for consistency. Dense model: active_b = params_b.

58.1±1.0#216 of 719 LLMsfrom 18 benchmark results across 7 sources

$ how it compares

The LLMs ranked either side of DeepSeek R1 Distill Llama 70B, out of 828 with a score, then more from DeepSeek.

#modelscorewhereRAMreleased
258Llama 3.3 Nemotron Super 49B v1 NVIDIA58±2local~22GB2025-03
259Olmo 3.1 32B Think Ai258±1local~21GB2025-12
260DeepSeek R1 Distill Llama 70B DeepSeek58±1api $0.80/M +local~30GB2025-01
261Phi-4-reasoning 14B Microsoft58±1api $0.22/M +local~10GB2025-04
262Olmo 3 32B Think Ai258±1local~21GB2025-11
more from DeepSeek
21DeepSeek V4.1 Flash DeepSeek78±1api $0.26/M +local~537GB2026-09
31DeepSeek V4 Pro (0813) DeepSeek76±1api $1.63/M +local~1010GB2026-08
35DeepSeek V3.2 Speciale DeepSeek76±3comparison only

see all LLMs · find the ones that fit your machine

$ what it's good at

coding▁▁▁▃▆██▇▆▄▃▁ #243 of 458 · 14% of its score · 5 tests

Beats Qwen3-VL 32B Instruct, Solar Open 100B (102B-A12B), GLM-4.7-Flash here, though they rank higher overall. Trails QwQ-32B-Preview, InternLM2.5 20B Chat, Mistral Small Instruct 22B (2409) here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
LiveCodeBench (version unspecified) lab57.5% ✎▄····▄▄··▄█▄ beats 5/6medium · avg 41%▇▇▇ 7.5%at its level
CodeReviewQA: change localisation (hard), invariant accuracy47.3% ✎█▂▁▁▂▂▂▃▅▅▅▇ beats 35/47medium · avg 36%▇ 2.8%sharp · at its level · noisy
CodeReviewQA: automated code refinement, exact match rate28.7% ✎▁▁▁▁▁▁▂▂▃█▅▄ beats 19/47hard · avg 33%▇ 2.0%at its level
CodeReviewQA: solution identification (hard), invariant accuracy68.8% ✎▃··▁▂▂▃▂█▃▄▂ beats 37/47medium · avg 54%▇ 1.7%sharp · noisy
CodeReviewQA: change type recognition, invariant accuracy61.2% ✎▁▁▁······▁▃█ beats 11/47easy · avg 67%▇ 0.0%blunt · at its level · noisy
math▁▂▃▅▆▅▄█▄▁▁▁ #166 of 398 · 37% of its score · 5 tests

Beats MiniMax-M2.1, Mistral Medium 3.5 (128B), GPT-4.1 here, though they rank higher overall. Trails Olmo 3 7B Think, Nemotron Nano 9B v2, Qwen3-VL 4B Thinking here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
MATH level 589.9% ✎▂▂▅▆▄█▆▃▅▃▃▄ beats 78/92hard · avg 32%▇▇▇▇▇▇▇ 17%sharp · at its level
OTIS mock AIME 2024–202551.4% ✎▃▃▄▃▃▃▇▇█▃▂▅ beats 78/193very hard · avg 8.5%▇▇▇ 7.7%sharp · at its level · noisy
AIME 2024 lab +170.0% ✎▃·▃·▃▃···█·▃ beats 5/7medium · avg 43%▇▇ 5.6%sharp · at its level · noisy · several runs
AIME 202551.5% ✎█··▄··█··▄·▄ beats 0/6hard · avg 34%▇▇ 4.3%sharp · at its level · noisy
MATH-500 lab94.5% ✎▄·▄▄·▄▄█···▄ beats 6/7very easy · avg 91%▇ 1.9%maxed out
other tests5 not in a task · 49% of its score
testresultvs the others in its rundifficultycountswhy
GPQA Diamond +255.7% ✎▁▁▁▁▂▄▆▄▅▇█▅ beats 79/208hard · avg 40%▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 40%sharp · at its level · several runs
BRIDGE clinical text, zero-shot average (87 tasks)39.8% ✎▁▁▁▁▁▂▂▃▄▆█▃ beats 70/97hard · avg 32%▇▇ 4.5%blunt · at its level
Judgemark v2 (LLM-as-judge for creative writing)59.7% ✎▃▃··▃▅▅···█▃ beats 3/10medium · avg 55%▇ 2.8%at its level
EuroEval HellaSwag (English), test split, 5-shot generative MC, accuracy39.8% ✎▃·▁▁▂▃▂▅█▇▄▁ beats 21/100medium · avg 68%▇ 1.1%noisy
Omni-MATH63.4% ✎█·█·█·██··██ beats 1/6medium · avg 64%▇ 0.5%blunt · at its level · noisy

$ where to run it

Through an API

api · Novita AI · $0.80/M · $0.8 in / $0.8 out per 1M tokens · 131k context · novita.ai ↗ · price sourceUSD 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.

llama.cpp · recommended

version

  1. Install llama.cpp · Homebrew and winget installs name the commands llama-cli and llama-server instead.
    macOS / Linux
    curl -LsSf https://llama.app/install.sh | sh
    Windows
    irm https://llama.app/install.ps1 | iex

    other ways to install ↗

  2. Run DeepSeek R1 Distill Llama 70B · needs ~74.98 GB · from huggingface.co/unsloth
    chat in the terminal
    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)
    llama serve -hf unsloth/DeepSeek-R1-Distill-Llama-70B-GGUF:Q8_0

LM Studio

version

  1. Install LM Studio · Open the app once before using the lms command.

    download it (macOS / Windows / Linux) ↗

  2. Run DeepSeek R1 Distill Llama 70B · needs ~39.7 GB, Mac only · from huggingface.co/mlx-community

    open in LM Studio ↗

  3. Or download it in the terminal
    lms get https://huggingface.co/mlx-community/DeepSeek-R1-Distill-Llama-70B-4bit -y
  4. Then run lms ls to see its name, and lms chat followed by that name

Ollama

  1. Install Ollama

    download it (macOS / Windows) ↗

    Linux
    curl -fsSL https://ollama.com/install.sh | sh
  2. Run DeepSeek R1 Distill Llama 70B · needs ~42.52 GB · from ollama.com
    chat in the terminal
    ollama run deepseek-r1:70b

MLX (Mac)

version

  1. Install MLX (Mac) · Apple silicon Macs only, text only.
    with uv
    uv tool install mlx-lm
    or
    pip install mlx-lm
  2. Run DeepSeek R1 Distill Llama 70B · needs ~39.7 GB, Mac only · from huggingface.co/mlx-community
    chat in the terminal
    mlx_lm.chat --model mlx-community/DeepSeek-R1-Distill-Llama-70B-4bit
  3. Or start an OpenAI-compatible server at http://localhost:8080/v1
    mlx_lm.server --model mlx-community/DeepSeek-R1-Distill-Llama-70B-4bit

Weights and documentation: huggingface.co/deepseek-ai ↗

$ how the score was worked out · Epoch AI — internal evaluatio… 40%, Microsoft — Phi-4-reasoning T… 23%, deepseek-ai/DeepSeek-R1 model… 22% and 4 more

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

sourceresultsshare of the scorelicence
Epoch AI — internal evaluations3▇▇▇▇▇▇▇▇▇▇▇▇ 40%CC BY 4.0
Microsoft — Phi-4-reasoning Technical Report4▇▇▇▇▇▇▇ 23%CC BY 4.0
deepseek-ai/DeepSeek-R1 model card · lab-reported4▇▇▇▇▇▇▇ 22%mit
CodeReviewQA: The Code Review Comprehension Assessment for Large Language Models (arXiv:2503.16167v2)4▇▇ 7%CC-BY-4.0
BRIDGE Medical Leaderboard (Harvard Medical School / Brigham and Women's YLab)1▇ 5%MIT
Judgemark v2 results (EQ-Bench)1▇ 3%MIT
EuroEval leaderboards (results collection), English generative leaderboard1▇ 1%MIT
prior · pull toward the middle▇ 0%
weighted average= 58.1
$ quantization evidence · 33 measurements

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.

precisiontestfull → quantizedkept
AWQ 4-bit (AutoAWQ; vLLM)Average of AIME 2024, FOLIO and Temporal accuracy (published 'Avg'; 3 passes each; vLLM, T=0.6, top-p 0.95, 32768 max tokens)0.818 → 0.80498.3%source ↗
AWQ 4-bit (AutoAWQ; vLLM)AIME 2024 accuracy (mean of 3 passes; vLLM, T=0.6, top-p 0.95, 32768 max tokens)0.656 → 0.63496.6%source ↗
FP8 W8A8 dynamic (FP8 weights symmetric per-channel, FP8 activations dynamic per-token; llm-compressor FP8_DYNAMIC)AIME 2024 pass@10.678 → 0.692102.0%source ↗
FP8 W8A8 dynamic (FP8 weights symmetric per-channel, FP8 activations dynamic per-token; llm-compressor FP8_DYNAMIC)Reasoning average of 3 tasks (AIME 2024 pass@1, MATH-500 pass@1, GPQA Diamond pass@1)0.762 → 0.765100.3%source ↗
FP8 W8A8 dynamic (FP8 weights symmetric per-channel, FP8 activations dynamic per-token; llm-compressor FP8_DYNAMIC)GSM8K 5-shot strict-match (lm-evaluation-harness, vLLM)0.93 → 0.93100.0%source ↗
FP8 W8A8 dynamic (FP8 weights symmetric per-channel, FP8 activations dynamic per-token; llm-compressor FP8_DYNAMIC)MATH-500 pass@10.953 → 0.95199.8%source ↗
FP8 W8A8 dynamic (FP8 weights symmetric per-channel, FP8 activations dynamic per-token; llm-compressor FP8_DYNAMIC)OpenLLM v1 average of 6 tasks (ARC-Challenge 25-shot, GSM8K 5-shot, HellaSwag 10-shot, MMLU 5-shot, TruthfulQA MC2 0-shot, Winogrande 5-shot; lm-evaluation-harness `openllm`, vLLM)0.757 → 0.75599.7%source ↗
FP8 W8A8 dynamic (FP8 weights symmetric per-channel, FP8 activations dynamic per-token; llm-compressor FP8_DYNAMIC)GPQA Diamond pass@10.656 → 0.65199.4%source ↗
FP8 W8A8 dynamic (FP8 weights symmetric per-channel, FP8 activations dynamic per-token; llm-compressor FP8_DYNAMIC)MMLU 5-shot acc (lm-evaluation-harness, vLLM)0.78 → 0.77499.2%source ↗
FP8 W8A8 dynamic (FP8 weights symmetric per-channel, FP8 activations dynamic per-token; llm-compressor FP8_DYNAMIC)IFEval 0-shot instruction-level strict acc (OpenLLM v2 config, chat template)0.424 → 0.42199.2%source ↗
FP8 W8A8 dynamic (FP8 weights symmetric per-channel, FP8 activations dynamic per-token; llm-compressor FP8_DYNAMIC)OpenLLM v2 average of 6 tasks (IFEval, BBH, MATH-Hard, GPQA, MuSR, MMLU-Pro; lm-evaluation-harness `leaderboard`, chat template, vLLM; values exactly as published, per-task normalisation not stated)0.154 → 0.14996.6%source ↗
GPTQ 3-bit (vLLM)Average of AIME 2024, FOLIO and Temporal accuracy (published 'Avg'; 3 passes each; vLLM, T=0.6, top-p 0.95, 32768 max tokens)0.818 → 0.72688.8%source ↗
GPTQ 3-bit (vLLM)AIME 2024 accuracy (mean of 3 passes; vLLM, T=0.6, top-p 0.95, 32768 max tokens)0.656 → 0.46771.2%source ↗
GPTQ 4-bit (vLLM)AIME 2024 accuracy (mean of 3 passes; vLLM, T=0.6, top-p 0.95, 32768 max tokens)0.656 → 0.667101.7%source ↗
GPTQ 4-bit (vLLM)Average of AIME 2024, FOLIO and Temporal accuracy (published 'Avg'; 3 passes each; vLLM, T=0.6, top-p 0.95, 32768 max tokens)0.818 → 0.81299.3%source ↗
INT4 W4A16 (weights only, symmetric per-group, group size 128; card states GPTQ via llm-compressor)OpenLLM v1 average of 6 tasks (ARC-Challenge 25-shot, GSM8K 5-shot, HellaSwag 10-shot, MMLU 5-shot, TruthfulQA MC2 0-shot, Winogrande 5-shot; lm-evaluation-harness `openllm`, vLLM)0.757 → 0.759100.2%source ↗
INT4 W4A16 (weights only, symmetric per-group, group size 128; card states GPTQ via llm-compressor)GSM8K 5-shot strict-match (lm-evaluation-harness, vLLM)0.93 → 0.93100.0%source ↗
INT4 W4A16 (weights only, symmetric per-group, group size 128; card states GPTQ via llm-compressor)IFEval 0-shot instruction-level strict acc (OpenLLM v2 config, chat template)0.424 → 0.42499.9%source ↗
INT4 W4A16 (weights only, symmetric per-group, group size 128; card states GPTQ via llm-compressor)MATH-500 pass@10.953 → 0.95299.9%source ↗
INT4 W4A16 (weights only, symmetric per-group, group size 128; card states GPTQ via llm-compressor)MMLU 5-shot acc (lm-evaluation-harness, vLLM)0.78 → 0.77198.9%source ↗
INT4 W4A16 (weights only, symmetric per-group, group size 128; card states GPTQ via llm-compressor)Reasoning average of 3 tasks (AIME 2024 pass@1, MATH-500 pass@1, GPQA Diamond pass@1)0.762 → 0.7598.3%source ↗
INT4 W4A16 (weights only, symmetric per-group, group size 128; card states GPTQ via llm-compressor)GPQA Diamond pass@10.656 → 0.6497.7%source ↗
INT4 W4A16 (weights only, symmetric per-group, group size 128; card states GPTQ via llm-compressor)AIME 2024 pass@10.678 → 0.65696.7%source ↗
INT4 W4A16 (weights only, symmetric per-group, group size 128; card states GPTQ via llm-compressor)OpenLLM v2 average of 6 tasks (IFEval, BBH, MATH-Hard, GPQA, MuSR, MMLU-Pro; lm-evaluation-harness `leaderboard`, chat template, vLLM; values exactly as published, per-task normalisation not stated)0.154 → 0.13688.1%source ↗
INT8 W8A8 (weights symmetric per-channel, activations symmetric per-token; SmoothQuant + GPTQ via llm-compressor)GSM8K 5-shot strict-match (lm-evaluation-harness, vLLM)0.93 → 0.936100.6%source ↗
INT8 W8A8 (weights symmetric per-channel, activations symmetric per-token; SmoothQuant + GPTQ via llm-compressor)OpenLLM v1 average of 6 tasks (ARC-Challenge 25-shot, GSM8K 5-shot, HellaSwag 10-shot, MMLU 5-shot, TruthfulQA MC2 0-shot, Winogrande 5-shot; lm-evaluation-harness `openllm`, vLLM)0.757 → 0.76100.4%source ↗
INT8 W8A8 (weights symmetric per-channel, activations symmetric per-token; SmoothQuant + GPTQ via llm-compressor)IFEval 0-shot instruction-level strict acc (OpenLLM v2 config, chat template)0.424 → 0.425100.1%source ↗
INT8 W8A8 (weights symmetric per-channel, activations symmetric per-token; SmoothQuant + GPTQ via llm-compressor)MATH-500 pass@10.953 → 0.953100.0%source ↗
INT8 W8A8 (weights symmetric per-channel, activations symmetric per-token; SmoothQuant + GPTQ via llm-compressor)MMLU 5-shot acc (lm-evaluation-harness, vLLM)0.78 → 0.78100.0%source ↗
INT8 W8A8 (weights symmetric per-channel, activations symmetric per-token; SmoothQuant + GPTQ via llm-compressor)AIME 2024 pass@10.678 → 0.67899.9%source ↗
INT8 W8A8 (weights symmetric per-channel, activations symmetric per-token; SmoothQuant + GPTQ via llm-compressor)Reasoning average of 3 tasks (AIME 2024 pass@1, MATH-500 pass@1, GPQA Diamond pass@1)0.762 → 0.7699.7%source ↗
INT8 W8A8 (weights symmetric per-channel, activations symmetric per-token; SmoothQuant + GPTQ via llm-compressor)GPQA Diamond pass@10.656 → 0.6599.1%source ↗
INT8 W8A8 (weights symmetric per-channel, activations symmetric per-token; SmoothQuant + GPTQ via llm-compressor)OpenLLM v2 average of 6 tasks (IFEval, BBH, MATH-Hard, GPQA, MuSR, MMLU-Pro; lm-evaluation-harness `leaderboard`, chat template, vLLM; values exactly as published, per-task normalisation not stated)0.154 → 0.15299.0%source ↗

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated · sources and licences · json