A curated starting point for running AI on your computer. Find a model that suits your task, compare memory estimates, and download the actual model file.
Free to download. Each model keeps its own license and usage terms.
WEIGHTS → RUNTIME → LOCAL CHAT
LOCAL AIGGUF
+YOUR MACHINEAIREADY FOR YOUR NEXT IDEA
CHOOSE→DOWNLOAD→RUN
↓
model.ggufReal weights. Local runtime.
+
27 reviewed downloadsGGUF · Free model weightsCatalog reviewed
01 / CHOOSE YOUR WEIGHTS
A good fit. A clear download.
Start with your task and available memory. These are reviewed selections, with specific file versions and source details you can inspect.
INSPECT THE DIFFERENCES
Compare selected models
Choose two or three models. Memory values are planning estimates at roughly 4,096 tokens. Runtime and architecture support vary; these values are not speed or quality benchmarks. On Apple silicon, RAM and GPU share one memory pool; do not add their capacities.
Apache-2.0: license and applicable notices must be preserved; changes must be identified when distributing modified copies. Review the linked terms.
Memory figures are estimates for one user at roughly 4,096 tokens. Runtime, context length and other applications change requirements. These are metadata-reviewed downloads, not benchmark results.
02GGUF · Q8_0
Qwen · publisher release
Qwen3 1.7B
Multilingual assistant with switchable thinking and direct-answer modes, supplied as a publisher-hosted GGUF.
Apache-2.0: license and applicable notices must be preserved; changes must be identified when distributing modified copies. Review the linked terms.
Memory figures are estimates for one user at roughly 4,096 tokens. Runtime, context length and other applications change requirements. These are metadata-reviewed downloads, not benchmark results.
03GGUF · Q4_K_M
Qwen · publisher release
Qwen3 4B
Multilingual assistant with switchable thinking and direct-answer modes, supplied as a publisher-hosted GGUF.
Apache-2.0: license and applicable notices must be preserved; changes must be identified when distributing modified copies. Review the linked terms.
Memory figures are estimates for one user at roughly 4,096 tokens. Runtime, context length and other applications change requirements. These are metadata-reviewed downloads, not benchmark results.
04GGUF · Q4_K_M
Qwen · publisher release
Qwen3 8B
Multilingual assistant with switchable thinking and direct-answer modes, supplied as a publisher-hosted GGUF.
Apache-2.0: license and applicable notices must be preserved; changes must be identified when distributing modified copies. Review the linked terms.
Memory figures are estimates for one user at roughly 4,096 tokens. Runtime, context length and other applications change requirements. These are metadata-reviewed downloads, not benchmark results.
05GGUF · Q4_K_M
Qwen · publisher release
Qwen3 14B
Multilingual assistant with switchable thinking and direct-answer modes, supplied as a publisher-hosted GGUF.
Apache-2.0: license and applicable notices must be preserved; changes must be identified when distributing modified copies. Review the linked terms.
Memory figures are estimates for one user at roughly 4,096 tokens. Runtime, context length and other applications change requirements. These are metadata-reviewed downloads, not benchmark results.
06GGUF · Q4_K_M
Qwen · publisher release
Qwen3 32B
Multilingual assistant with switchable thinking and direct-answer modes, supplied as a publisher-hosted GGUF.
Apache-2.0: license and applicable notices must be preserved; changes must be identified when distributing modified copies. Review the linked terms.
Memory figures are estimates for one user at roughly 4,096 tokens. Runtime, context length and other applications change requirements. These are metadata-reviewed downloads, not benchmark results.
07GGUF · Q4_K_M
Qwen · publisher release
Qwen2.5 0.5B Instruct
Compact instruction-tuned assistant for local chat and multilingual text tasks.
Apache-2.0: license and applicable notices must be preserved; changes must be identified when distributing modified copies. Review the linked terms.
Memory figures are estimates for one user at roughly 4,096 tokens. Runtime, context length and other applications change requirements. These are metadata-reviewed downloads, not benchmark results.
08GGUF · Q4_K_M
Qwen · publisher release
Qwen2.5 1.5B Instruct
Compact instruction-tuned assistant for local chat and multilingual text tasks.
Apache-2.0: license and applicable notices must be preserved; changes must be identified when distributing modified copies. Review the linked terms.
Memory figures are estimates for one user at roughly 4,096 tokens. Runtime, context length and other applications change requirements. These are metadata-reviewed downloads, not benchmark results.
09GGUF · Q4_K_M
Qwen · publisher release
Qwen2.5 Coder 0.5B Instruct
Small instruction-tuned coding model for local code generation, explanation, and repair experiments.
Apache-2.0: license and applicable notices must be preserved; changes must be identified when distributing modified copies. Review the linked terms.
Memory figures are estimates for one user at roughly 4,096 tokens. Runtime, context length and other applications change requirements. These are metadata-reviewed downloads, not benchmark results.
10GGUF · Q4_K_M
Qwen · publisher release
Qwen2.5 Coder 1.5B Instruct
Small instruction-tuned coding model for local code generation, explanation, and repair experiments.
Apache-2.0: license and applicable notices must be preserved; changes must be identified when distributing modified copies. Review the linked terms.
Memory figures are estimates for one user at roughly 4,096 tokens. Runtime, context length and other applications change requirements. These are metadata-reviewed downloads, not benchmark results.
11GGUF · Q8_0
Hugging Face · publisher release
SmolLM2 360M Instruct
Small English instruction model for experimenting with local assistants on modest hardware.
Apache-2.0: license and applicable notices must be preserved; changes must be identified when distributing modified copies. Review the linked terms.
Memory figures are estimates for one user at roughly 4,096 tokens. Runtime, context length and other applications change requirements. These are metadata-reviewed downloads, not benchmark results.
12GGUF · Q4_K_M
Hugging Face · publisher release
SmolLM2 1.7B Instruct
Small English instruction model for experimenting with local assistants on modest hardware.
Apache-2.0: license and applicable notices must be preserved; changes must be identified when distributing modified copies. Review the linked terms.
Memory figures are estimates for one user at roughly 4,096 tokens. Runtime, context length and other applications change requirements. These are metadata-reviewed downloads, not benchmark results.
13GGUF · Q4
Microsoft · publisher release
Phi 3 mini 4k instruct
Microsoft instruction model for English chat, reasoning, and coding; this artifact uses the publisher’s Q4 label.
MIT: retain the copyright and permission notice. Review the linked publisher license for the exact terms.
Memory figures are estimates for one user at roughly 4,096 tokens. Runtime, context length and other applications change requirements. These are metadata-reviewed downloads, not benchmark results.
14GGUF · Q4_K_M
IBM · publisher release
granite 3.3 2b instruct
IBM instruction model with structured thinking support and multilingual dialogue capabilities.
Apache-2.0: license and applicable notices must be preserved; changes must be identified when distributing modified copies. Review the linked terms.
Memory figures are estimates for one user at roughly 4,096 tokens. Runtime, context length and other applications change requirements. These are metadata-reviewed downloads, not benchmark results.
15GGUF · Q4_K_M
IBM · publisher release
granite 3.3 8b instruct
IBM instruction model with structured thinking support and multilingual dialogue capabilities.
Apache-2.0: license and applicable notices must be preserved; changes must be identified when distributing modified copies. Review the linked terms.
Memory figures are estimates for one user at roughly 4,096 tokens. Runtime, context length and other applications change requirements. These are metadata-reviewed downloads, not benchmark results.
16GGUF · Q4_K_M
DeepSeek · conversion by bartowski
DeepSeek R1 Distill Qwen 1.5B
Qwen-based distilled reasoning model for math and coding, converted to GGUF by Bartowski.
DeepSeek releases these distilled weights under MIT and explicitly identifies Qwen-derived models as originally Apache-2.0. Preserve applicable MIT and upstream Apache notices. Community GGUF conversion.
Memory figures are estimates for one user at roughly 4,096 tokens. Runtime, context length and other applications change requirements. These are metadata-reviewed downloads, not benchmark results.
17GGUF · Q4_K_M
DeepSeek · conversion by bartowski
DeepSeek R1 Distill Qwen 7B
Qwen-based distilled reasoning model for math and coding, converted to GGUF by Bartowski.
DeepSeek releases these distilled weights under MIT and explicitly identifies Qwen-derived models as originally Apache-2.0. Preserve applicable MIT and upstream Apache notices. Community GGUF conversion.
Memory figures are estimates for one user at roughly 4,096 tokens. Runtime, context length and other applications change requirements. These are metadata-reviewed downloads, not benchmark results.
18GGUF · Q4_K_M
DeepSeek · conversion by bartowski
DeepSeek R1 Distill Qwen 14B
Qwen-based distilled reasoning model for math and coding, converted to GGUF by Bartowski.
DeepSeek releases these distilled weights under MIT and explicitly identifies Qwen-derived models as originally Apache-2.0. Preserve applicable MIT and upstream Apache notices. Community GGUF conversion.
Memory figures are estimates for one user at roughly 4,096 tokens. Runtime, context length and other applications change requirements. These are metadata-reviewed downloads, not benchmark results.
19GGUF · Q4_K_M
DeepSeek · conversion by bartowski
DeepSeek R1 Distill Qwen 32B
Qwen-based distilled reasoning model for math and coding, converted to GGUF by Bartowski.
DeepSeek releases these distilled weights under MIT and explicitly identifies Qwen-derived models as originally Apache-2.0. Preserve applicable MIT and upstream Apache notices. Community GGUF conversion.
Memory figures are estimates for one user at roughly 4,096 tokens. Runtime, context length and other applications change requirements. These are metadata-reviewed downloads, not benchmark results.
20GGUF · Q4_K_M
Mistral AI · conversion by bartowski
Mistral 7B Instruct v0.3
General instruction model with function-calling support, converted to GGUF by Bartowski.
Apache-2.0: license and applicable notices must be preserved; changes must be identified when distributing modified copies. Review the linked terms.
Memory figures are estimates for one user at roughly 4,096 tokens. Runtime, context length and other applications change requirements. These are metadata-reviewed downloads, not benchmark results.
21GGUF · Q4_K_M
Mistral AI / NVIDIA · conversion by bartowski
Mistral Nemo Instruct 2407
Multilingual instruction model with code training, converted to a single GGUF by Bartowski.
Apache-2.0: license and applicable notices must be preserved; changes must be identified when distributing modified copies. Review the linked terms.
Memory figures are estimates for one user at roughly 4,096 tokens. Runtime, context length and other applications change requirements. These are metadata-reviewed downloads, not benchmark results.
22GGUF · Q4_K_M
TinyLlama · conversion by TheBloke
TinyLlama 1.1B Chat v1.0
Compact chat model for introductory local inference, converted to GGUF by TheBloke.
Apache-2.0: license and applicable notices must be preserved; changes must be identified when distributing modified copies. Review the linked terms.
Memory figures are estimates for one user at roughly 4,096 tokens. Runtime, context length and other applications change requirements. These are metadata-reviewed downloads, not benchmark results.
23GGUF · Q4_K_M
IBM · publisher release
Granite 4.0 Micro
Compact IBM instruction model for multilingual chat and tool-calling workflows, supplied in the publisher’s GGUF repository.
Apache-2.0: preserve the license and applicable notices, identify modifications when distributing modified copies, and comply with the linked terms.
Runtime compatibility: Use a current llama.cpp or LM Studio build with Granite 4 GGUF support. Micro is the dense transformer variant; longer contexts require more memory. Runtime inference has not been tested here.
Memory figures are estimates for one user at roughly 4,096 tokens. Runtime, context length and other applications change requirements. These are metadata-reviewed downloads, not benchmark results.
24GGUF · Q4_K_M
Hugging Face · conversion by ggml-org
SmolLM3 3B
Small multilingual assistant with thinking and direct-answer modes, converted to GGUF by the llama.cpp project’s ggml-org account.
Apache-2.0: preserve the license and applicable notices, identify modifications when distributing modified copies, and comply with the linked terms.
Runtime compatibility: Use a current llama.cpp or compatible app with the bundled SmolLM3 chat template. The GGUF card requires --jinja to enable thinking in llama.cpp; use /think or /no_think through the system prompt. Runtime inference has not been tested here.
Memory figures are estimates for one user at roughly 4,096 tokens. Runtime, context length and other applications change requirements. These are metadata-reviewed downloads, not benchmark results.
25GGUF · Q4_K_M
Qwen · conversion by bartowski / LM Studio Community
Qwen3 Coder 30B A3B Instruct
Mixture-of-experts coding assistant with 30.5B total and 3.3B active parameters, quantized by Bartowski for LM Studio Community.
Apache-2.0: preserve the license and applicable notices, identify modifications when distributing modified copies, and comply with the linked terms.
Runtime compatibility: Use a current llama.cpp or LM Studio build with Qwen3 MoE support and the bundled chat template. This is a non-thinking model. Its 3.3B active parameters do not reduce the stored weights to a 3B-sized model; long contexts need additional memory. Runtime inference has not been tested here.
Memory figures are estimates for one user at roughly 4,096 tokens. Runtime, context length and other applications change requirements. These are metadata-reviewed downloads, not benchmark results.
26GGUF · Q4_K_M
IBM · publisher release
Granite Switch 4.1 8B
IBM's 8B-base assistant with embedded Decision and Guardian adapters, supplied as an official GGUF for text workflows. Requires a runtime with Granite Switch support.
Apache-2.0: preserve the license and applicable notices, identify modifications when distributing modified copies, and comply with the linked terms.
Runtime compatibility: Requires a recent llama.cpp build with graniteswitch support and the bundled Jinja chat template. This download is listed for text inference; audio support and runtime performance have not been tested here. Compatibility with other apps has not been verified.
Memory figures are estimates for one user at roughly 4,096 tokens. Runtime, context length and other applications change requirements. These are metadata-reviewed downloads, not benchmark results.
27GGUF · Q4_K_M
IBM · publisher release
Granite Switch 4.1 30B
IBM's 30B-base assistant with embedded Decision and Guardian adapters, supplied as an official GGUF for text workflows. Requires a runtime with Granite Switch support.
Apache-2.0: preserve the license and applicable notices, identify modifications when distributing modified copies, and comply with the linked terms.
Runtime compatibility: Requires a recent llama.cpp build with graniteswitch support and the bundled Jinja chat template. This download is listed for text inference; audio support and runtime performance have not been tested here. Compatibility with other apps has not been verified.
Memory figures are estimates for one user at roughly 4,096 tokens. Runtime, context length and other applications change requirements. These are metadata-reviewed downloads, not benchmark results.
⌕
No models match these filters.
Try a broader search or a different RAM capacity. This curated library is a starting selection.
↗
The file starts here. The bytes come from the original host.
Download buttons point to actual, version-pinned model files on Hugging Face. Your browser handles the transfer; your network and the host determine download speed. We do not rehost model weights.
02 / FROM FILE TO FIRST PROMPT
Downloaded. Now make it run.
A model file needs a compatible runtime. Start with a desktop app, or import your downloaded GGUF into Ollama.
01DESKTOP APP
Start with LM Studio
A visual route to loading a model and chatting locally.
Install LM Studio ↗ for your computer and launch it once to initialize its CLI.
Download a GGUF from a model card above.
Import it with the terminal command below, then select the model in LM Studio.
Load the model and start a chat. Reduce context length if memory is tight.
lms import "/path/to/model.gguf" --copy
Replace the example path with the downloaded file. The --copy option retains the original file and uses additional disk space. On Windows, use a quoted path such as "C:\Models\model.gguf". Use a recent runtime version that supports your model’s architecture.
A focused launch selection, built around transparent files and practical choices.
What does “free” mean here?
There is no charge from this website to download the listed weights. Model licenses still govern commercial use, redistribution, and other conditions. Read the linked license and upstream model card for your intended use.
Will a model run on my computer?
RAM and VRAM figures are estimates for the selected quantization, not guarantees or speed benchmarks. Context length, runtime support, GPU offload, and available memory matter. On Apple silicon, RAM and GPU use the same unified memory pool; do not add their capacities together.
GGUF is a file format used by local runtimes. Q4 and Q5 labels indicate quantization: smaller files trade some precision for lower memory use. A bigger model is not automatically a better fit for your task.
How do I inspect or verify a download?
Open “Source, integrity & requirements” on a model card for the repository, pinned revision, exact filename, and SHA-256 checksum. Compare the downloaded file’s SHA-256 with that value. Matching hashes confirm the downloaded bytes match the listed file; they do not evaluate a model’s behavior.
On Windows: Get-FileHash "C:\Models\model.gguf" -Algorithm SHA256. On macOS: shasum -a 256 model.gguf. On Linux: sha256sum model.gguf.
How is this library kept current?
Each card shows its availability check date, and each download is pinned to a specific reviewed revision. A new upstream version may exist without being listed here yet. New models and changed licenses need review before joining this selection.
Availability checks confirm file metadata at the time shown. They do not continuously monitor your download or confirm runtime compatibility on every device.
Can I use a model offline?
Once the model and a compatible runtime are installed, local inference can run without sending prompts to a hosted model. Your runtime may also offer cloud services, plugins, or telemetry; check its settings for the behavior you want.
