utBaidu · Downloadable
ERNIE-4.5-21B-A3B-Paddle
Generates text for conversations and writing.
- Source listed
- Jun 28, 2025
- Parameters
- 21B
- 4-bit weights only
- ≈ 10.5 GB
- License
- apache-2.0
YOUR MEMORY BUDGET
Local AI models for 16GB VRAM.
Explore models by theoretical weight storage. This shortlist covers available downloadable language models with a known or name-derived parameter count.
14 GBleft for weights at 4-bit
16 GB budget − 2 GB reserve
16 GB budget − 2 GB reserve
The reserve is your assumption, not a measured overhead allowance. The default 2 GB may be too little for your runtime or context length. Results are not confirmed fits and have no verified speed rating. Weight estimates use decimal GB and do not include quantization overhead. For timed results on one 12GB machine, see measured speeds.
1,030 models within the weight budget
Largest parameter count first · not a quality rankingBaidu · Downloadable
ERNIE-4.5-21B-A3B-PT
Generates text for conversations and writing.
- Source listed
- Jun 28, 2025
- Parameters
- 21B
- 4-bit weights only
- ≈ 10.5 GB
- License
- apache-2.0
Baidu · Downloadable
ERNIE-4.5-21B-A3B-Thinking
Generates text for conversations and writing. Supports a reasoning mode.
- Source listed
- Sep 8, 2025
- Parameters
- 21B
- 4-bit weights only
- ≈ 10.5 GB
- License
- apache-2.0
Microsoft · Downloadable
OptiMind-SFT
Generates text for conversations and writing.
- Source listed
- Dec 10, 2025
- Parameters
- 20.9B
- 4-bit weights only
- ≈ 10.5 GB
- License
- mit
M-A-P · Downloadable
340M-20B-DeltaNet-hybrid-12-1
Generates text, according to its published model architecture (DeltaNetForCausalLM). No task is listed on Hugging Face, so check the model card for intended use.
- Source listed
- Mar 4, 2025
- Parameters
- 20B
- 4-bit weights only
- ≈ 10 GB
- License
- Not disclosed
M-A-P · Downloadable
340M-20B-DeltaNet-hybrid-24-1
Generates text, according to its published model architecture (DeltaNetForCausalLM). No task is listed on Hugging Face, so check the model card for intended use.
- Source listed
- Mar 4, 2025
- Parameters
- 20B
- 4-bit weights only
- ≈ 10 GB
- License
- Not disclosed
M-A-P · Downloadable
340M-20B-DeltaNet-hybrid-3-1
Generates text, according to its published model architecture (DeltaNetForCausalLM). No task is listed on Hugging Face, so check the model card for intended use.
- Source listed
- Mar 4, 2025
- Parameters
- 20B
- 4-bit weights only
- ≈ 10 GB
- License
- Not disclosed
M-A-P · Downloadable
340M-20B-DeltaNet-hybrid-6-1
Generates text, according to its published model architecture (DeltaNetForCausalLM). No task is listed on Hugging Face, so check the model card for intended use.
- Source listed
- Mar 4, 2025
- Parameters
- 20B
- 4-bit weights only
- ≈ 10 GB
- License
- Not disclosed
M-A-P · Downloadable
340M-20B-DeltaNet-pure
Generates text, according to its published model architecture (DeltaNetForCausalLM). No task is listed on Hugging Face, so check the model card for intended use.
- Source listed
- Mar 4, 2025
- Parameters
- 20B
- 4-bit weights only
- ≈ 10 GB
- License
- Not disclosed
M-A-P · Downloadable
340M-20B-GatedDeltaNet-hybrid-12-1
Generates text, according to its published model architecture (GatedDeltaNetForCausalLM). No task is listed on Hugging Face, so check the model card for intended use.
- Source listed
- May 10, 2025
- Parameters
- 20B
- 4-bit weights only
- ≈ 10 GB
- License
- Not disclosed
M-A-P · Downloadable
340M-20B-GatedDeltaNet-hybrid-3-1
Generates text, according to its published model architecture (GatedDeltaNetForCausalLM). No task is listed on Hugging Face, so check the model card for intended use.
- Source listed
- Mar 3, 2025
- Parameters
- 20B
- 4-bit weights only
- ≈ 10 GB
- License
- Not disclosed
M-A-P · Downloadable
340M-20B-GatedDeltaNet-hybrid-6-1
Generates text, according to its published model architecture (GatedDeltaNetForCausalLM). No task is listed on Hugging Face, so check the model card for intended use.
- Source listed
- Mar 3, 2025
- Parameters
- 20B
- 4-bit weights only
- ≈ 10 GB
- License
- Not disclosed
M-A-P · Downloadable
340M-20B-GatedDeltaNet-New-hybrid-24-1
Generates text, according to its published model architecture (GatedDeltaNetForCausalLM). No task is listed on Hugging Face, so check the model card for intended use.
- Source listed
- Jun 9, 2025
- Parameters
- 20B
- 4-bit weights only
- ≈ 10 GB
- License
- Not disclosed
M-A-P · Downloadable
340M-20B-GatedDeltaNet-pure-baseline
Generates text, according to its published model architecture (GatedDeltaNetForCausalLM). No task is listed on Hugging Face, so check the model card for intended use.
- Source listed
- Jun 7, 2025
- Parameters
- 20B
- 4-bit weights only
- ≈ 10 GB
- License
- Not disclosed
M-A-P · Downloadable
340M-20B-GLA-hybrid-12-1
Generates text, according to its published model architecture (GLAForCausalLM). No task is listed on Hugging Face, so check the model card for intended use.
- Source listed
- Mar 3, 2025
- Parameters
- 20B
- 4-bit weights only
- ≈ 10 GB
- License
- Not disclosed
M-A-P · Downloadable
340M-20B-GLA-hybrid-24-1
Generates text, according to its published model architecture (GLAForCausalLM). No task is listed on Hugging Face, so check the model card for intended use.
- Source listed
- Mar 3, 2025
- Parameters
- 20B
- 4-bit weights only
- ≈ 10 GB
- License
- Not disclosed
M-A-P · Downloadable
340M-20B-GLA-hybrid-3-1
Generates text, according to its published model architecture (GLAForCausalLM). No task is listed on Hugging Face, so check the model card for intended use.
- Source listed
- Mar 3, 2025
- Parameters
- 20B
- 4-bit weights only
- ≈ 10 GB
- License
- Not disclosed
M-A-P · Downloadable
340M-20B-GLA-hybrid-6-1
Generates text, according to its published model architecture (GLAForCausalLM). No task is listed on Hugging Face, so check the model card for intended use.
- Source listed
- Mar 3, 2025
- Parameters
- 20B
- 4-bit weights only
- ≈ 10 GB
- License
- Not disclosed
M-A-P · Downloadable
340M-20B-GLA-pure-baseline
Generates text, according to its published model architecture (GLAForCausalLM). No task is listed on Hugging Face, so check the model card for intended use.
- Source listed
- May 10, 2025
- Parameters
- 20B
- 4-bit weights only
- ≈ 10 GB
- License
- Not disclosed
M-A-P · Downloadable
340M-20B-HGRN-hybrid-12-1
Generates text, according to its published model architecture (HGRNForCausalLM). No task is listed on Hugging Face, so check the model card for intended use.
- Source listed
- Apr 20, 2025
- Parameters
- 20B
- 4-bit weights only
- ≈ 10 GB
- License
- Not disclosed
M-A-P · Downloadable
340M-20B-HGRN-hybrid-24-1
Generates text, according to its published model architecture (HGRNForCausalLM). No task is listed on Hugging Face, so check the model card for intended use.
- Source listed
- Apr 20, 2025
- Parameters
- 20B
- 4-bit weights only
- ≈ 10 GB
- License
- Not disclosed
M-A-P · Downloadable
340M-20B-HGRN-hybrid-3-1
Generates text, according to its published model architecture (HGRNForCausalLM). No task is listed on Hugging Face, so check the model card for intended use.
- Source listed
- Apr 20, 2025
- Parameters
- 20B
- 4-bit weights only
- ≈ 10 GB
- License
- Not disclosed
M-A-P · Downloadable
340M-20B-HGRN-hybrid-6-1
Generates text, according to its published model architecture (HGRNForCausalLM). No task is listed on Hugging Face, so check the model card for intended use.
- Source listed
- Apr 20, 2025
- Parameters
- 20B
- 4-bit weights only
- ≈ 10 GB
- License
- Not disclosed
M-A-P · Downloadable
340M-20B-HGRN-pure-baseline
Generates text, according to its published model architecture (HGRNForCausalLM). No task is listed on Hugging Face, so check the model card for intended use.
- Source listed
- Apr 20, 2025
- Parameters
- 20B
- 4-bit weights only
- ≈ 10 GB
- License
- Not disclosed
What changes with a 16GB budget?
With 2 GB set aside, 14 GB remains for weights. At 4-bit that allows up to 28 billion parameters in the theoretical calculation. An 8-billion-parameter model alone needs approximately 4 GB at this precision, before its conversation cache and runtime.
Why might a listed model still fail to load?
Its actual weight format may need more space, the runtime may not support its architecture, or the conversation cache may exceed your reserve. Vision models can also need space for image processing. Check the original model card and measure your exact setup before relying on it.
Read the memory guide and technical sources · Compare two models