utMiniMax · Downloadable
SynLogic-32B
Generates text for conversations and writing.
- Source listed
- May 30, 2025
- Parameters
- 32B
- 4-bit weights only
- ≈ 16 GB
- License
- mit
YOUR MEMORY BUDGET
Local AI models for 24GB VRAM.
Explore models by theoretical weight storage. This shortlist covers available downloadable language models with a known or name-derived parameter count.
22 GBleft for weights at 4-bit
24 GB budget − 2 GB reserve
24 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.
1,129 models within the weight budget
Largest parameter count first · not a quality rankingMiniMax · Downloadable
SynLogic-Mix-3-32B
Generates text for conversations and writing.
- Source listed
- May 30, 2025
- Parameters
- 32B
- 4-bit weights only
- ≈ 16 GB
- License
- mit
NVIDIA · Downloadable
Cosmos-Reason2-32B
Understands text and images and responds in text. Supports a reasoning mode.
- Source listed
- Apr 29, 2026
- Parameters
- 32B
- 4-bit weights only
- ≈ 16 GB
- License
- other
OpenMOSS · Downloadable
Qwen3-32B-ABC
Generates text for conversations and writing.
- Source listed
- Jan 14, 2026
- Parameters
- 32B
- 4-bit weights only
- ≈ 16 GB
- License
- apache-2.0
Qwen · Alibaba · Downloadable
Qwen2.5-Coder-32B-Instruct
Designed for coding and software tasks. Generates text for conversations and writing.
- Source listed
- Nov 6, 2024
- Parameters
- 32B
- 4-bit weights only
- ≈ 16 GB
- License
- apache-2.0
Qwen · Alibaba · Downloadable
Qwen3-32B
Generates text for conversations and writing.
- Source listed
- Apr 27, 2025
- Parameters
- 32B
- 4-bit weights only
- ≈ 16 GB
- License
- apache-2.0
Qwen · Alibaba · Downloadable
Qwen3-VL-32B-Instruct
Understands text and images and responds in text.
- Source listed
- Oct 19, 2025
- Parameters
- 32B
- 4-bit weights only
- ≈ 16 GB
- License
- apache-2.0
Qwen · Alibaba · Downloadable
Qwen3-VL-32B-Thinking
Understands text and images and responds in text. Supports a reasoning mode.
- Source listed
- Oct 19, 2025
- Parameters
- 32B
- 4-bit weights only
- ≈ 16 GB
- License
- apache-2.0
Qwen · Alibaba · Downloadable
WebWorld-32B
Generates text for conversations and writing.
- Source listed
- Feb 13, 2026
- Parameters
- 32B
- 4-bit weights only
- ≈ 16 GB
- License
- apache-2.0
StepFun · Downloadable
Qwen2.5-32B-DialogueReason
Generates text, according to its published model architecture (Qwen2ForCausalLM). No task is listed on Hugging Face, so check the model card for intended use.
- Source listed
- May 9, 2025
- Parameters
- 32B
- 4-bit weights only
- ≈ 16 GB
- License
- apache-2.0
StepFun · Downloadable
StepFun-Formalizer-32B
Generates text for conversations and writing.
- Source listed
- Aug 19, 2025
- Parameters
- 32B
- 4-bit weights only
- ≈ 16 GB
- License
- apache-2.0
Z.ai · Downloadable
GLM-4-32B-0414
Generates text for conversations and writing.
- Source listed
- Apr 7, 2025
- Parameters
- 32B
- 4-bit weights only
- ≈ 16 GB
- License
- mit
Z.ai · Downloadable
GLM-4-32B-Base-0414
Generates text for conversations and writing.
- Source listed
- Apr 7, 2025
- Parameters
- 32B
- 4-bit weights only
- ≈ 16 GB
- License
- mit
Z.ai · Downloadable
GLM-Z1-32B-0414
Generates text for conversations and writing.
- Source listed
- Apr 8, 2025
- Parameters
- 32B
- 4-bit weights only
- ≈ 16 GB
- License
- mit
Z.ai · Downloadable
GLM-Z1-Rumination-32B-0414
Generates text for conversations and writing.
- Source listed
- Apr 13, 2025
- Parameters
- 32B
- 4-bit weights only
- ≈ 16 GB
- License
- mit
Z.ai · Downloadable
SWE-Dev-32B
Generates text for conversations and writing.
- Source listed
- Apr 6, 2025
- Parameters
- 32B
- 4-bit weights only
- ≈ 16 GB
- License
- mit
Z.ai · Downloadable
GLM-4.7-Flash
Generates text for conversations and writing.
- Source listed
- Jan 19, 2026
- Parameters
- 31.2B
- 4-bit weights only
- ≈ 15.6 GB
- License
- mit
Google · Downloadable
gemma-4-31B
Understands text and images and responds in text.
- Source listed
- Mar 12, 2026
- Parameters
- 31B
- 4-bit weights only
- ≈ 15.5 GB
- License
- apache-2.0
Google · Downloadable
gemma-4-31B-it
Understands text and images and responds in text.
- Source listed
- Mar 11, 2026
- Parameters
- 31B
- 4-bit weights only
- ≈ 15.5 GB
- License
- apache-2.0
NVIDIA · Downloadable
Ising-Calibration-1.5-31B
Understands text and images and responds in text.
- Source listed
- Jul 13, 2026
- Parameters
- 31B
- 4-bit weights only
- ≈ 15.5 GB
- License
- other
Perplexity · Downloadable
browsesafe
Generates text for conversations and writing.
- Source listed
- Nov 14, 2025
- Parameters
- 30.5B
- 4-bit weights only
- ≈ 15.3 GB
- License
- mit
Tencent · Downloadable
Hy-Embodied-VLM-1.0
Understands text and images and responds in text.
- Source listed
- Jul 14, 2026
- Parameters
- 30.5B
- 4-bit weights only
- ≈ 15.3 GB
- License
- apache-2.0
inclusionAI · Downloadable
AReaL-tau2-airline-sft-30B
Generates text, according to its published model architecture (Qwen3MoeForCausalLM). No task is listed on Hugging Face, so check the model card for intended use.
- Source listed
- Mar 4, 2026
- Parameters
- 30B
- 4-bit weights only
- ≈ 15 GB
- License
- apache-2.0
inclusionAI · Downloadable
AReaL-tau2-retail-sft-30B
Generates text, according to its published model architecture (Qwen3MoeForCausalLM). No task is listed on Hugging Face, so check the model card for intended use.
- Source listed
- Mar 4, 2026
- Parameters
- 30B
- 4-bit weights only
- ≈ 15 GB
- License
- apache-2.0
What changes with a 24GB budget?
With 2 GB set aside, 22 GB remains for weights. At 4-bit that allows up to 44 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