Model database
Every open model, and what it takes to run it
How much memory each model needs, the cheapest GPU that runs it well, and where it shines.
How to read this page
Memory figures are for the recommended quantization (usually Q4_K_M) with 8K tokens of context. Mixture-of-experts models show total and active parameters. Click any model for VRAM at every quantization, speed on popular GPUs, and setup commands.
Small models (under 10B)
Fast on any modern GPU with 8–12GB. Good for chat, summarizing, and autocomplete.
Mid-size models (10–40B)
The local sweet spot. 16–24GB GPUs run these at good quality.
Alibaba Qwen
Qwen3.8 27B
27B · 256K ctx
Alibaba Qwen
Qwen3.6 35B-A3B
35B (3B active) · 256K ctx
OpenAI
gpt-oss-20b
20.9B (3.6B active) · 128K ctx
Gemma 4 31B
31B · 256K ctx
Gemma 4 26B A4B
26B (3.8B active) · 256K ctx
Microsoft
Phi-4
14.7B · 16K ctx
Large models (40–150B)
48GB+ of memory, multiple GPUs, or a high-memory Mac. MoE models here run faster than their size suggests.
OpenAI
gpt-oss-120b
117B (5.1B active) · 128K ctx
Alibaba Qwen
Qwen3.5 122B-A10B
122B (10B active) · 256K ctx
Z.ai
GLM-4.5-Air
106B (12B active) · 128K ctx
Meta
Llama 4 Scout
109B (17B active) · 10.5M ctx
Frontier-scale models (150B+)
Open models that rival the best closed ones. They need 128GB+ of memory or a multi-GPU server.
Reference
All models
Click a column to sort.
| Model | Params | Memory | Cheapest way to run it | Context | License | Released |
|---|---|---|---|---|---|---|
| GLM-5.3Z.ai | 744B (40B active) | ~466GBQ4_K_M | M3 Ultra 512GB | 1.0M | GLM-5.3 custom license | Aug 28, 2026 |
| GLM-5.3-FlashZ.ai | 320B (18B active) | ~201GBQ4_K_M | 8× RTX 5090 | 1.0M | MIT license | Aug 26, 2026 |
| Qwen3.8 27BAlibaba Qwen | 27B | ~18GBQ4_K_M | RTX 4060 Ti 16GB | 256K | Apache 2.0 license | Aug 14, 2026 |
| DeepSeek V4 FlashDeepSeek | 284B (13B active) | ~168GBFP4 | 8× RTX 3090 | 1.0M | MIT license | Apr 24, 2026 |
| Qwen3.6 35B-A3BAlibaba Qwen | 35B (3B active) | ~24GBQ4_K_M | RTX 4000 Ada | 256K | Apache 2.0 license | Apr 16, 2026 |
| Gemma 4 31BGoogle | 31B | ~22GBQ4_K_M | RTX 4000 Ada | 256K | Apache 2.0 license | Apr 2, 2026 |
| Gemma 4 26B A4BGoogle | 26B (3.8B active) | ~17GBQ4_K_M | RTX 4060 Ti 16GB | 256K | Apache 2.0 license | Apr 2, 2026 |
| Gemma 4 E4BGoogle | 8B | ~6.3GBQ4_K_M | RTX 5060 | 128K | Apache 2.0 license | Apr 2, 2026 |
| Qwen3.5 9BAlibaba Qwen | 9B | ~7.2GBQ4_K_M | RTX 5060 | 256K | Apache 2.0 license | Mar 2, 2026 |
| Qwen3.5 122B-A10BAlibaba Qwen | 122B (10B active) | ~77GBQ4_K_M | RTX PRO 5000 72GB | 256K | Apache 2.0 license | Feb 24, 2026 |
| gpt-oss-20bOpenAI | 20.9B (3.6B active) | ~15GBMXFP4 | RTX 4060 Ti 16GB | 128K | Apache 2.0 license | Aug 5, 2025 |
| gpt-oss-120bOpenAI | 117B (5.1B active) | ~68GBMXFP4 | RTX PRO 5000 72GB | 128K | Apache 2.0 license | Aug 5, 2025 |
| Qwen3 Coder 30B-A3B InstructAlibaba QwenOlder | 30.5B (3.3B active) | ~21GBQ4_K_M | RTX 4000 Ada | 256K | Apache 2.0 license | Jul 31, 2025 |
| Qwen3 30B-A3B (2507)Alibaba QwenOlder | 30.5B (3.3B active) | ~21GBQ4_K_M | RTX 4000 Ada | 256K | Apache 2.0 license | Jul 29, 2025 |
| GLM-4.5-AirZ.ai | 106B (12B active) | ~68GBQ4_K_M | RTX PRO 5000 72GB | 128K | MIT license | Jul 28, 2025 |
| Qwen3 235B-A22B (2507)Alibaba QwenOlder | 235B (22B active) | ~149GBQ4_K_M | 8× RTX 3090 | 256K | Apache 2.0 license | Jul 21, 2025 |
| Mistral Small 3.2 24B InstructMistral AIOlder | 24B | ~17GBQ4_K_M | RTX 4060 Ti 16GB | 128K | Apache 2.0 license | Jun 20, 2025 |
| DeepSeek R1 (0528)DeepSeekOlder | 671B (37B active) | ~420GBQ4_K_M | M3 Ultra 512GB | 128K | MIT license | May 28, 2025 |
| Qwen3 32BAlibaba QwenOlder | 32.8B | ~23GBQ4_K_M | RTX 4000 Ada | 32K | Apache 2.0 license | Apr 29, 2025 |
| Qwen3 8BAlibaba QwenOlder | 8.2B | ~7GBQ4_K_M | RTX 5060 | 32K | Apache 2.0 license | Apr 29, 2025 |
| Qwen3 14BAlibaba QwenOlder | 14.8B | ~11GBQ4_K_M | RTX 3060 12GB | 32K | Apache 2.0 license | Apr 29, 2025 |
| Llama 4 ScoutMeta | 109B (17B active) | ~70GBQ4_K_M | RTX PRO 5000 72GB | 10.5M | Llama 4 Community License | Apr 5, 2025 |
| Gemma 3 27BGoogleOlder | 27.4B | ~19GBQ4_K_M | RTX 4060 Ti 16GB | 128K | Gemma Terms of Use | Mar 12, 2025 |
| Gemma 3 12BGoogleOlder | 12.2B | ~9GBQ4_K_M | RTX 5060 | 128K | Gemma Terms of Use | Mar 12, 2025 |
| Phi-4Microsoft | 14.7B | ~12GBQ4_K_M | RTX 3060 12GB | 16K | MIT license | Dec 12, 2024 |
| Llama 3.3 70B InstructMetaOlder | 70.6B | ~47GBQ4_K_M | RTX A6000 | 128K | Llama 3.3 Community License | Dec 6, 2024 |
| Llama 3.1 8B InstructMetaOlder | 8B | ~6.7GBQ4_K_M | RTX 5060 | 128K | Llama 3.1 Community License | Jul 23, 2024 |
| Llama 2 7B ChatMetaOlder | 6.7B | ~6.9GBQ4_K_M | RTX 5060 | 4K | Llama 2 Community License | Jul 18, 2023 |
Models marked "Older" have been superseded by newer releases but are kept for reference because many people still run them.
Frequently asked questions
What is the best local LLM right now?
The most capable open-weight model in our catalog is GLM-5.3, but it needs serious hardware. For a single consumer GPU, Qwen3.8 27B is the strongest model that fits in 24GB. The best model for you depends on your memory: check the best-models page for your GPU.
What does "MoE" mean?
Mixture of experts. The model has many expert sub-networks but only uses a few for each token, so a 30B MoE model with 3B active parameters needs the memory of a 30B model but generates text at close to the speed of a 3B model.
What does Q4_K_M mean?
It is a 4-bit GGUF quantization format from llama.cpp. It shrinks model weights to roughly 30% of their 16-bit size with only a small loss in quality, which is why it is the default for local use.