compute.info

Models / Open weights

Which open models fit your hardware?

Creator-published model specifications from Hugging Face, paired with transparent memory estimates for common accelerator capacities.

Creator model cards

Open-model directory

10 of 10 models

DeepSeek R1

DeepSeek R1

Full reasoning model, not a distilled checkpoint; 37B parameters are activated per token but all 671B weights must be stored.

textMIT
Parameters
671B
Context
131K

DeepSeek V3

DeepSeek V3

Mixture-of-experts model with 37B parameters activated per token; local weight memory reflects 671B total parameters.

textMIT
Parameters
671B
Context
131K

Gemma 3

Gemma 3 1B IT

Text-only model with gated license acknowledgment.

textGemma license
Parameters
1B
Context
33K

Gemma 3

Gemma 3 4B IT

Multimodal input; images are normalized to 896×896 and encoded as 256 tokens.

textimageGemma license
Parameters
4B
Context
131K

Llama 3.2

Llama 3.2 11B Vision

Image-plus-text is officially supported only in English; custom license restrictions apply.

textimageLlama 3.2 Community License
Parameters
10.6B
Context
131K

Llama 3.2

Llama 3.2 1B Instruct

Custom commercial license with gated access and acceptable-use policy; actual architecture count is approximately 1.23B.

textLlama 3.2 Community License
Parameters
1.23B
Context
131K

Llama 3.3

Llama 3.3 70B Instruct

Custom commercial license, gated access, and acceptable-use policy apply.

textLlama 3.3 Community License
Parameters
70B
Context
131K

Mistral Small

Mistral Small 3.1 24B

Supports image understanding, not image generation; requires Mistral-specific serving and tokenizer settings.

textimageApache-2.0
Parameters
24B
Context
131K

Qwen3

Qwen3 0.6B

Repository name uses 0.6B; Hugging Face safetensors metadata reports approximately 752M parameters.

textApache-2.0
Parameters
0.752B
Context
33K

Qwen3

Qwen3 30B-A3B

Mixture-of-experts model with 3.3B parameters activated per token; weight memory reflects 30.5B total parameters. Validated to 131,072 tokens with YaRN.

textApache-2.0
Parameters
30.5B
Context
33K

Estimated local deployment

How model size maps to accelerator memory

Estimated weights plus 20% loading overhead. Runtime, KV cache, context length, and software can require substantially more memory.

Estimated 4-bit weight memoryGB, including 20% overhead · logarithmic bar scale
Qwen3 0.6B0.6 GB
Qwen3 30B-A3B22.9 GB
Gemma 3 1B IT0.8 GB
Gemma 3 4B IT3.0 GB
Mistral Small 3.1 24B18.0 GB
Llama 3.2 1B Instruct0.9 GB
Llama 3.2 11B Vision7.9 GB
Llama 3.3 70B Instruct52.5 GB
DeepSeek V3503.3 GB
DeepSeek R1503.3 GB

Method

Weights are only the starting point.

The estimate uses parameters × planning bits ÷ 8 × 1.20: 5 planning bits for Q4 and 9 for Q8 to allow for scales, metadata, and mixed-precision tensors. “Comfortable” reserves at least 20% of device memory. “Tight” means estimated weights fit but leave less headroom.

This is a capacity screen—not a speed benchmark or guarantee that a particular runtime, context length, multimodal projector, or GPU configuration will work.

Estimated 4-bit model weight fit by available accelerator memory
ModelParametersQ4 estimateQ8 estimate8 GB12 GB16 GB24 GB32 GB48 GB80 GB141 GB192 GB256 GB512 GB
Qwen3 0.6B0.752B0.6 GB1.0 GBComfortableComfortableComfortableComfortableComfortableComfortableComfortableComfortableComfortableComfortableComfortable
Qwen3 30B-A3B30.5B22.9 GB41.2 GBNoNoNoTightComfortableComfortableComfortableComfortableComfortableComfortableComfortable
Gemma 3 1B IT1B0.8 GB1.3 GBComfortableComfortableComfortableComfortableComfortableComfortableComfortableComfortableComfortableComfortableComfortable
Gemma 3 4B IT4B3.0 GB5.4 GBComfortableComfortableComfortableComfortableComfortableComfortableComfortableComfortableComfortableComfortableComfortable
Mistral Small 3.1 24B24B18.0 GB32.4 GBNoNoNoComfortableComfortableComfortableComfortableComfortableComfortableComfortableComfortable
Llama 3.2 1B Instruct1.23B0.9 GB1.7 GBComfortableComfortableComfortableComfortableComfortableComfortableComfortableComfortableComfortableComfortableComfortable
Llama 3.2 11B Vision10.6B7.9 GB14.3 GBTightComfortableComfortableComfortableComfortableComfortableComfortableComfortableComfortableComfortableComfortable
Llama 3.3 70B Instruct70B52.5 GB94.5 GBNoNoNoNoNoNoComfortableComfortableComfortableComfortableComfortable
DeepSeek V3671B503.3 GB905.9 GBNoNoNoNoNoNoNoNoNoNoTight
DeepSeek R1671B503.3 GB905.9 GBNoNoNoNoNoNoNoNoNoNoTight

What comes next

Capacity first, performance next.

This release answers whether model weights are likely to fit. A useful throughput comparison must additionally pin the model revision, quantization, prompt and generation lengths, batch size, runtime, kernels, driver, and exact hardware configuration.

The next hardware dataset will add source-grounded tokens-per-second measurements without mixing incomparable benchmark setups.