Open Source LLM Download Center

Top 10 open-source LLMs with download links, install guides & hardware recommendations

🚀 Quick Start: Install Ollama and run any model in 3 steps

No account needed, data stays local · MacBook / Mac mini / Windows / Linux supported

brew install ollama

curl -fsSL https://ollama.com/install.sh | sh

irm https://ollama.com/install.ps1 | iex

Install

ollama pull qwen3

Download

ollama run qwen3

Start Chat

Filter by Use Case

Featured Open Source Models

DeepSeek-R1

DeepSeek AI

MIT

671B

Params

16GB+

Min VRAM

★★★★★

Reasoning

Top reasoning model globally, math & code on par with proprietary flagships, efficient MoE architecture

Daily Chat Code Dev Document

ollama pull deepseek-r1

↓ Download 📖 Tutorial

Qwen 3.5

Alibaba

Apache 2.0

0.6B~235B

Params

4GB+

Min VRAM

★★★★★

Chinese NLP

Best Chinese comprehension globally, 0.6B–235B scales covering low-end to flagship hardware

Daily Chat Code Dev Document Low-end

ollama pull qwen3

↓ Download 📖 Tutorial

MiniMax M2

MiniMax

CC-BY-NC

456B

Params

24GB+

Min VRAM

★★★★☆

Infer Speed

Sparse MoE ultra-fast inference, low power, pair with OpenClaw for 24/7 local AI assistant

Daily Chat Document

huggingface-cli download MiniMaxAI/MiniMax-M1-40k

↓ Download 📖 Tutorial

智谱 GLM

Zhipu AI

Apache 2.0

9B~32B

Params

8GB+

Min VRAM

★★★★☆

Code Ability

Tsinghua flagship open model, leading in code & tool calling, mature deployment ecosystem

Daily Chat Code Dev Document

ollama pull glm4

↓ Download 📖 Tutorial

Llama 4

Mistral Large 2

Mistral AI

MRL 2.0

123B

Params

24GB+

Min VRAM

★★★★☆

Multilingual

Europe's top open model, excellent code & multilingual, privacy-compliant

Daily Chat Code Dev

ollama pull mistral-large

↓ Download 🔗 Official Docs

Gemma 3

Google

Gemma ToU

1B~27B

Params

4GB+

Min VRAM

★★★★☆

Lightweight

Google lightweight open model, runs 1B on 4GB VRAM, first choice for low-end devices

Daily Chat Low-end

ollama pull gemma3

↓ Download 🔗 Official Docs

Phi-4

Microsoft

MIT

14B

Params

8GB

Min VRAM

★★★★★

Value

Microsoft precision small model, 14B outperforms same-size rivals, MIT license for commercial use

Code Dev Low-end

ollama pull phi4

↓ Download 🔗 Official Docs

MiniCPM

Tsinghua / ModelBest

Apache 2.0

3B~4B

Params

4GB

Min VRAM

★★★★☆

Edge Deploy

Ultra-lightweight model that runs on phones, best for edge deployment, excellent Chinese support

Daily Chat Low-end

ollama pull minicpm-v

↓ Download 🔗 Official Docs

Yi-34B

01.AI

Apache 2.0

34B

Params

16GB+

Min VRAM

★★★★☆

CN+EN

01.AI flagship open model, balanced Chinese-English, excellent long-doc understanding & writing

Daily Chat Document

ollama pull yi:34b

↓ Download 🔗 Official Docs

Hardware Reference

Choose models based on your GPU VRAM. More VRAM = larger model parameters you can run.

GPU VRAM	Recommended Models	Typical Device	Notes
4 GB	Gemma 3 (1B/4B) MiniCPM-3B Qwen3-0.6B	GTX 1650 / Laptop iGPU	Entry-level, daily chat & simple tasks
8 GB	Qwen3-8B Phi-4 (14B量化) GLM-9B	RTX 3060 / RTX 4060	Mainstream, smooth for code & document tasks
16 GB	DeepSeek-R1-14B Qwen3-14B Yi-34B量化	RTX 3080 / RTX 4070	Advanced, significantly better reasoning quality
24 GB+	DeepSeek-R1-32B Qwen3-32B Mistral-22B	RTX 3090 / RTX 4090	Flagship, approaches proprietary model quality

Tip: All models can be downloaded and run via Ollama: ollama pull <model-name>

Model Tier List

March 2026 · Unified benchmark comparison · Full bar = highest score

	模型	MMLU-Pro General Knowledge	GPQA Diamond Science Reasoning	SWE-Bench Code Repair	Arena Elo Human Preference	VRAM
S Top Tier — On par with leading proprietary models
S	Qwen 3.5 397B/17B · 阿里巴巴	84.6%	82.1%	62.5%	1451	8GB+
S	DeepSeek-R1 685B/37B · 深度求索	84.0%	85.3%	49.2%	1420	16GB+
S	智谱 GLM-5 744B/40B · 智谱AI	70.4%	86.0%	77.8%	1452	24GB+
A A Tier — Flagship capability, manageable hardware
A	Llama 4 Maverick 400B/17B · Meta	83.2%	78.5%	55.8%	1320	8GB+
A	Mistral Large 3 675B/41B · Mistral AI	82.8%	79.3%	54.1%	1315	24GB+
B B Tier — Runs on consumer GPU, excellent overall capability
B	Llama 4 Scout 109B/17B · Meta · 10M上下文	78.5%	74.2%	48.5%	1280	8GB
B	Gemma 3 27B 27B · Google	67.5%	42.4%	35.2%	1220	16GB
C C Tier — 4–8GB VRAM, best for edge/low-end devices
C	Phi-4 14B · Microsoft · MIT	75.2%	56.1%	41.3%	1200	8GB
C	Yi-1.5-34B 34B · 零一万物	63.1%	40.2%	31.5%	1140	16GB
C	MiniCPM-o 4.5 9B · 清华/面壁 · 多模态	58.3%	38.5%	28.1%	1150	6GB

MMLU-Pro General Knowledge

GPQA Diamond PhD-level Science

SWE-Bench Code Repair (max=77.8%)

Arena Elo Human Preference (max=1500)

Sources: Artificial Analysis · LMSYS Chatbot Arena · Official model reports (March 2026) · Some scores are community test estimates

Top Tier Comparable to top proprietary models, high hardware demand

智谱 GLM-5 Apache 2.0

744B 总 / 40B 激活 · 24GB+ Elo 1452

SWE-Bench

77.8%

GPQA

86.0%

MMLU-Pro

70.4%

Code #1 China Chip

DeepSeek-R1 MIT

685B 总 / 37B 激活 · 16GB+ MATH 97.3%

AIME 2025

79.8%

GPQA

85.3%

MMLU-Pro

84.0%

Best Reasoning Math #1

Qwen 3.5 Apache 2.0

397B 总 / 17B 激活 · 8GB+ Elo 1451

SWE-Bench

62.5%

GPQA

82.1%

MMLU-Pro

84.6%

Chinese #1 201 Languages

A Tier Flagship capability, manageable hardware requirements

Llama 4 Maverick Llama 4

400B 总 / 17B 激活 · 8GB+

SWE-Bench

55.8%

GPQA

78.5%

MMLU-Pro

83.2%

1M Context Largest Ecosystem

Mistral Large 3 Apache 2.0

675B 总 / 41B 激活 · 24GB+

SWE-Bench

54.1%

GPQA

79.3%

MMLU-Pro

82.8%

EU Compliant Strong Multilingual

B Tier Runs on single consumer GPU, excellent overall capability

Llama 4 Scout Llama 4

109B 总 / 17B 激活 · 单H100

MMLU-Pro

78.5%

Context

10M

10M Long Context Single GPU Run

Gemma 3 27B Gemma ToU

27B · 16GB 消费显卡

HumanEval

78.5%

MMLU-Pro

67.5%

GPQA

42.4%

Best Consumer GPU Multimodal

Download LLMs Fast – Speed Matters

Downloading 50GB+ model files from Hugging Face requires a stable high-speed global network

¥9/月

Low Monthly Fee

1000Mbps

Gigabit Bandwidth

70+国家

全球节点

30 Days

Money-back Guarantee

Free Trial VPN07 View Pricing

Open Source LLM Download Center

Filter by Use Case

Featured Open Source Models

DeepSeek-R1

Qwen 3.5

MiniMax M2

智谱 GLM

Llama 4

Mistral Large 2

Gemma 3

Phi-4

MiniCPM

Yi-34B

Hardware Reference

Model Tier List

Download LLMs Fast – Speed Matters

Related Tutorials

DeepSeek R1 – Advanced Reasoning at 1/10th the Cost

Qwen3.5 Ollama Setup: Run 0.8B–35B Free on PC & Mac

Qwen3.5 on iPhone: Run 9B AI Offline with MLX

Qwen3.5 Android: Best Apps to Run AI Locally