Model

DeepSeek-R1-Distill-Llama-70B

OPEN

DeepSeek · DeepSeek-R1 family · released Jan 20, 2025

Largest R1 distill: Llama-3.3-70B fine-tuned on R1 reasoning traces, the most popular self-hostable strong reasoner from the release.

ReasoningCodingVisionFunction callingTool useAgentic

1384.0

Elo · rank #139

Parameters

70B

Active params

70B (dense)

Context

128K tokens

Architecture

Dense Transformer (Llama-3.3-70B-Instruct architecture, GQA)

License

Llama 3.3 Community License

Languages

—

API price (in/out)

No hosted API

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

AIMEMath70.0%#105

best: GPT-5.2 · 100.0%

CodeforcesCoding1633#34

best: DeepSeek-V4-Pro (Think Max) · 3206

GPQA DiamondReasoning65.2%#144

best: GPT-5.6 · 94.6%

Humanity's Last ExamReasoning6.0%#110

best: Claude Sonnet 5 · 57.4%

LiveCodeBenchCoding57.5%#100

best: DeepSeek-V4-Pro (Think Max) · 93.5%

MATH-500Math94.5%#34

best: GPT-5 · 99.4%

Run it locally

VRAM @ Q4

42 GB

VRAM @ FP16

140 GB

Fits on (Q4)

M3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

Q4_K_M needs ~40-42GB, exceeding a single 24GB 4090; community reports ~12-15 tok/s on dual RTX 3090 (48GB) setups.

Quantizations

GGUF Q4_K_M · AWQ · GPTQ

DeepSeek-R1 family

Elo progression across releases

DeepSeek-R1 671BJan 20251574.5 DeepSeek-R1-Distill-Llama-70BJan 20251384.0 DeepSeek-R1-Distill-Qwen-32BJan 20251009.0 DeepSeek-R1-ZeroJan 20251435.8 DeepSeek-R1-0528May 20251757.4

Resources

HFdeepseek-ai/DeepSeek-R1-Distill-Llama-70B GHdeepseek-ai/DeepSeek-R1

API price weights · each benchmark row carries its own source badge (see methodology)