Can I run it?

One model. One machine. One honest answer.

These are curated combinations, not a generated matrix of every possible name. Each page uses the same published-weight, KV-cache, and runtime-overhead formula, then shows the context ceiling and the command to try.

Run the interactive calculator or inspect the formula.

too-large

gpt-oss-20b on NVIDIA GeForce RTX 3060 · 12 GB

13.8 GB estimated · MXFP4 · 0 token ceiling

a compact MoE package for local coding and agent experiments

tight

gpt-oss-20b on NVIDIA GeForce RTX 4060 Ti · 16 GB

13.8 GB estimated · MXFP4 · 95,740 token ceiling

a compact MoE package for local coding and agent experiments

comfortable

gpt-oss-20b on NVIDIA GeForce RTX 4090 · 24 GB

13.8 GB estimated · MXFP4 · 131,072 token ceiling

a compact MoE package for local coding and agent experiments

comfortable

gpt-oss-20b on NVIDIA GeForce RTX 5090 · 32 GB

13.8 GB estimated · MXFP4 · 131,072 token ceiling

a compact MoE package for local coding and agent experiments

tight

gpt-oss-20b on Mac mini M4 · 16 GB

13.8 GB estimated · MXFP4 · 95,740 token ceiling

a compact MoE package for local coding and agent experiments

too-large

Qwen3.6 35B-A3B on NVIDIA GeForce RTX 3060 · 12 GB

24.3 GB estimated · MXFP4 · 0 token ceiling

a larger MoE model that stores 35B parameters but activates about 3B per token

too-large

Qwen3.6 35B-A3B on NVIDIA GeForce RTX 4060 Ti · 16 GB

24.3 GB estimated · MXFP4 · 0 token ceiling

a larger MoE model that stores 35B parameters but activates about 3B per token

too-large

Qwen3.6 35B-A3B on NVIDIA GeForce RTX 4090 · 24 GB

24.3 GB estimated · MXFP4 · 0 token ceiling

a larger MoE model that stores 35B parameters but activates about 3B per token

comfortable

Qwen3.6 35B-A3B on NVIDIA GeForce RTX 5090 · 32 GB

24.7 GB estimated · Q4_K_M · 262,144 token ceiling

a larger MoE model that stores 35B parameters but activates about 3B per token

too-large

Qwen3.6 35B-A3B on Mac mini M4 · 16 GB

24.3 GB estimated · MXFP4 · 0 token ceiling

a larger MoE model that stores 35B parameters but activates about 3B per token

too-large

Qwen3.8 27B on NVIDIA GeForce RTX 3060 · 12 GB

19.4 GB estimated · Q4_K_M · 0 token ceiling

a dense 27B model where the full weight set, not a smaller expert path, determines capacity

too-large

Qwen3.8 27B on NVIDIA GeForce RTX 4060 Ti · 16 GB

19.4 GB estimated · Q4_K_M · 0 token ceiling

a dense 27B model where the full weight set, not a smaller expert path, determines capacity

comfortable

Qwen3.8 27B on NVIDIA GeForce RTX 4090 · 24 GB

19.4 GB estimated · Q4_K_M · 77,937 token ceiling

a dense 27B model where the full weight set, not a smaller expert path, determines capacity

comfortable

Qwen3.8 27B on NVIDIA GeForce RTX 5090 · 32 GB

19.4 GB estimated · Q4_K_M · 200,001 token ceiling

a dense 27B model where the full weight set, not a smaller expert path, determines capacity

too-large

Qwen3.8 27B on Mac mini M4 · 16 GB

19.4 GB estimated · Q4_K_M · 0 token ceiling

a dense 27B model where the full weight set, not a smaller expert path, determines capacity

comfortable

Gemma 4 12B Unified on NVIDIA GeForce RTX 3060 · 12 GB

9.2 GB estimated · Q4_K_M · 65,157 token ceiling

a mid-sized multimodal model for general local assistance

comfortable

Gemma 4 12B Unified on NVIDIA GeForce RTX 4060 Ti · 16 GB

9.2 GB estimated · Q4_K_M · 146,528 token ceiling

a mid-sized multimodal model for general local assistance

comfortable

Gemma 4 12B Unified on NVIDIA GeForce RTX 4090 · 24 GB

14.7 GB estimated · Q8_0 · 196,408 token ceiling

a mid-sized multimodal model for general local assistance

comfortable

Gemma 4 12B Unified on NVIDIA GeForce RTX 5090 · 32 GB

14.7 GB estimated · Q8_0 · 262,144 token ceiling

a mid-sized multimodal model for general local assistance

comfortable

Gemma 4 12B Unified on Mac mini M4 · 16 GB

9.2 GB estimated · Q4_K_M · 146,528 token ceiling

a mid-sized multimodal model for general local assistance

comfortable

Gemma 4 E4B on NVIDIA GeForce RTX 3060 · 12 GB

7.7 GB estimated · Q4_K_M · 131,072 token ceiling

an efficient Gemma variant whose stored embedding tables are larger than its effective parameter count suggests

comfortable

Gemma 4 E4B on NVIDIA GeForce RTX 4060 Ti · 16 GB

11.0 GB estimated · Q8_0 · 131,072 token ceiling

an efficient Gemma variant whose stored embedding tables are larger than its effective parameter count suggests

comfortable

Gemma 4 E4B on NVIDIA GeForce RTX 4090 · 24 GB

11.0 GB estimated · Q8_0 · 131,072 token ceiling

an efficient Gemma variant whose stored embedding tables are larger than its effective parameter count suggests

comfortable

Gemma 4 E4B on NVIDIA GeForce RTX 5090 · 32 GB

11.0 GB estimated · Q8_0 · 131,072 token ceiling

an efficient Gemma variant whose stored embedding tables are larger than its effective parameter count suggests

comfortable

Gemma 4 E4B on Mac mini M4 · 16 GB

11.0 GB estimated · Q8_0 · 131,072 token ceiling

an efficient Gemma variant whose stored embedding tables are larger than its effective parameter count suggests

too-large

Gemma 4 31B on NVIDIA GeForce RTX 3060 · 12 GB

22.2 GB estimated · Q4_K_M · 0 token ceiling

a workstation-class dense model with a much larger memory footprint

too-large

Gemma 4 31B on NVIDIA GeForce RTX 4060 Ti · 16 GB

22.2 GB estimated · Q4_K_M · 0 token ceiling

a workstation-class dense model with a much larger memory footprint

tight

Gemma 4 31B on NVIDIA GeForce RTX 4090 · 24 GB

22.2 GB estimated · Q4_K_M · 20,759 token ceiling

a workstation-class dense model with a much larger memory footprint

comfortable

Gemma 4 31B on NVIDIA GeForce RTX 5090 · 32 GB

22.2 GB estimated · Q4_K_M · 76,562 token ceiling

a workstation-class dense model with a much larger memory footprint

too-large

Gemma 4 31B on Mac mini M4 · 16 GB

22.2 GB estimated · Q4_K_M · 0 token ceiling

a workstation-class dense model with a much larger memory footprint

too-large

gpt-oss-120b on NVIDIA GeForce RTX 3060 · 12 GB

65.2 GB estimated · MXFP4 · 0 token ceiling

a large MoE model that can be offloaded only with a substantial system-memory budget

too-large

gpt-oss-120b on NVIDIA GeForce RTX 4060 Ti · 16 GB

65.2 GB estimated · MXFP4 · 0 token ceiling

a large MoE model that can be offloaded only with a substantial system-memory budget

too-large

gpt-oss-120b on NVIDIA GeForce RTX 4090 · 24 GB

65.2 GB estimated · MXFP4 · 0 token ceiling

a large MoE model that can be offloaded only with a substantial system-memory budget

too-large

gpt-oss-120b on NVIDIA GeForce RTX 5090 · 32 GB

65.2 GB estimated · MXFP4 · 0 token ceiling

a large MoE model that can be offloaded only with a substantial system-memory budget

too-large

gpt-oss-120b on Mac mini M4 · 16 GB

65.2 GB estimated · MXFP4 · 0 token ceiling

a large MoE model that can be offloaded only with a substantial system-memory budget

comfortable

Qwen3 8B on NVIDIA GeForce RTX 3060 · 12 GB

7.9 GB estimated · Q4_K_M · 32,768 token ceiling

a compact dense model for everyday chat and extraction

comfortable

Qwen3 8B on NVIDIA GeForce RTX 4060 Ti · 16 GB

11.6 GB estimated · Q8_0 · 32,768 token ceiling

a compact dense model for everyday chat and extraction

comfortable

Qwen3 8B on NVIDIA GeForce RTX 4090 · 24 GB

11.6 GB estimated · Q8_0 · 32,768 token ceiling

a compact dense model for everyday chat and extraction

comfortable

Qwen3 8B on NVIDIA GeForce RTX 5090 · 32 GB

11.6 GB estimated · Q8_0 · 32,768 token ceiling

a compact dense model for everyday chat and extraction

comfortable

Qwen3 8B on Mac mini M4 · 16 GB

11.6 GB estimated · Q8_0 · 32,768 token ceiling

a compact dense model for everyday chat and extraction

comfortable

Mistral 7B on NVIDIA GeForce RTX 3060 · 12 GB

7.0 GB estimated · Q4_K_M · 32,768 token ceiling

a mature 7B dense baseline that is useful for comparing older and newer local stacks

comfortable

Mistral 7B on NVIDIA GeForce RTX 4060 Ti · 16 GB

10.3 GB estimated · Q8_0 · 32,768 token ceiling

a mature 7B dense baseline that is useful for comparing older and newer local stacks

comfortable

Mistral 7B on NVIDIA GeForce RTX 4090 · 24 GB

10.3 GB estimated · Q8_0 · 32,768 token ceiling

a mature 7B dense baseline that is useful for comparing older and newer local stacks

comfortable

Mistral 7B on NVIDIA GeForce RTX 5090 · 32 GB

10.3 GB estimated · Q8_0 · 32,768 token ceiling

a mature 7B dense baseline that is useful for comparing older and newer local stacks

comfortable

Mistral 7B on Mac mini M4 · 16 GB

10.3 GB estimated · Q8_0 · 32,768 token ceiling

a mature 7B dense baseline that is useful for comparing older and newer local stacks

comfortable

DeepSeek R1 8B on NVIDIA GeForce RTX 3060 · 12 GB

7.5 GB estimated · Q4_K_M · 42,726 token ceiling

an 8B reasoning distillation whose long outputs can make context planning matter

comfortable

DeepSeek R1 8B on NVIDIA GeForce RTX 4060 Ti · 16 GB

11.1 GB estimated · Q8_0 · 45,778 token ceiling

an 8B reasoning distillation whose long outputs can make context planning matter

comfortable

DeepSeek R1 8B on NVIDIA GeForce RTX 4090 · 24 GB

11.1 GB estimated · Q8_0 · 106,815 token ceiling

an 8B reasoning distillation whose long outputs can make context planning matter

comfortable

DeepSeek R1 8B on NVIDIA GeForce RTX 5090 · 32 GB

11.1 GB estimated · Q8_0 · 131,072 token ceiling

an 8B reasoning distillation whose long outputs can make context planning matter

comfortable

DeepSeek R1 8B on Mac mini M4 · 16 GB

11.1 GB estimated · Q8_0 · 45,778 token ceiling

an 8B reasoning distillation whose long outputs can make context planning matter