Best Local AI Models for RTX 4090 and 64GB RAM
Calculated local AI model fits for an RTX 4090 24GB workstation with 64GB RAM.
Yes. A computer with 64GB RAM can run small local AI models. A safe first choice from the current MamiLens list is GPT-OSS 20B. This is a memory-fit estimate, not a speed promise.
The device we checked
MamiLens keeps 8GB for the operating system and background work. That leaves a working model budget of about 56GB.
Best models to start with
GPT-OSS 20B
GPT-OSS 20B is a curated mixture-of-experts profile for reasoning, coding, tools workloads.
- Calculated weights
- 14 GB
- Estimated total RAM
- 8 GB
- RAM left
- 56 GB
- Expected path
- Verified GPU
- Speed
- 25 to 47 tokens/s estimate
- Evidence
- Calculated estimate
ollama run gpt-oss:20bOpen evidence-labelled model profile →Qwen 2.5 14B
Qwen 2.5 14B is a curated dense profile for general, writing, coding workloads.
- Calculated weights
- 9 GB
- Estimated total RAM
- 9 GB
- RAM left
- 55 GB
- Expected path
- Verified GPU
- Speed
- 39 to 73 tokens/s estimate
- Evidence
- Calculated estimate
ollama run qwen2.5:14bOpen evidence-labelled model profile →Phi-4 Reasoning 14B
A reasoning-focused Phi-4 variant for difficult math, science and coding tasks.
- Calculated weights
- 11 GB
- Estimated total RAM
- 8 GB
- RAM left
- 56 GB
- Expected path
- Verified GPU
- Speed
- 32 to 60 tokens/s estimate
- Evidence
- Calculated estimate
ollama run phi4-reasoning:14bOpen evidence-labelled model profile →Qwen 3 14B
Qwen 3 14B is a curated dense profile for general, coding, reasoning workloads.
- Calculated weights
- 9.3 GB
- Estimated total RAM
- 8 GB
- RAM left
- 56 GB
- Expected path
- Verified GPU
- Speed
- 37 to 71 tokens/s estimate
- Evidence
- Calculated estimate
ollama run qwen3:14bOpen evidence-labelled model profile →DeepSeek R1 14B
DeepSeek R1 14B is a curated dense profile for reasoning, math, coding workloads.
- Calculated weights
- 9 GB
- Estimated total RAM
- 9 GB
- RAM left
- 55 GB
- Expected path
- Verified GPU
- Speed
- 39 to 73 tokens/s estimate
- Evidence
- Calculated estimate
ollama run deepseek-r1:14bOpen evidence-labelled model profile →Recommended inference engine
Ollama or llama.cpp with CUDA
This is the simplest supported starting path for NVIDIA acceleration.
More CPU cores can help, but memory bandwidth and engine settings still matter.How accurate is this result?
This page uses the same calculator rules as the live MamiLens tool. It includes model weights, runtime overhead, operating-system reserve, context reserve and simultaneous-user cache when selected. A displayed speed range is calculated from published memory bandwidth. It is not a benchmark from this exact computer.
Start with the first model. Test a normal task, record response speed and watch memory use before using it for important work.