What's My AIHardware fit page built from reference hardware bands and catalog-backed model fit.
Hardware fitWorkstation desktop

what local ai

What local AI can I run on a 48 GB workstation?

Desktop and workstation owners who want a believable starting point for large local reasoning and coding models. This page uses the maintained catalog and a calibrated hardware band to answer the common hardware-search version of the question without pretending that a public shared-device cluster already exists.

reference bandWorkstation desktop
starter modelLlama 3.3 70B
memory guide48 GB
tier guide70B

benchmark first

Use the benchmark before you trust the reference band.

This band is where 70B-class local AI becomes realistic, but even here the benchmark still matters before frontier-scale downloads. The benchmark turns this reference guide into a machine-specific answer before you spend time downloading models that are too large for the actual browser-visible hardware.

Benchmark this device for this hardware bandCompare known machines

starter models

Best first models for this hardware band

Llama 3.3 70B

70B class • 40.0 GB minimum

Llama 3.3 70B remains the most established American large-model jump for a serious local machine that can already hold a 70B-class package.

Open model page

gpt-oss-20b

34B class • 15.5 GB minimum

gpt-oss-20b is the clearest midrange American local-model pick when you want a serious reasoning assistant without jumping straight into a 32B-class package.

Open model page

OLMo 3.1 Instruct 32B

34B class • 19.5 GB minimum

OLMo 3.1 32B is the strongest Apache-licensed American 32B-class option, but it asks for more memory than gpt-oss-20b to reach a clean first run.

Open model page

runtime paths

Pick the runtime after you confirm the size band

Runtime choice comes second here. Use the benchmark to confirm the model size band, then use the runtime pages for the cleanest first pull inside Ollama, LM Studio, or llama.cpp.

P0Static

Runtime page

Best local models for Ollama

Search intent: ollama best model

Best for the quickest path from benchmark result to a real local run.

Runtime guide + catalog coverage

Open page
P1Static

Runtime page

Best local models for llama.cpp

Search intent: llama.cpp best model

Best for people who care about low-level control, serving flags, and GGUF tuning.

Runtime guide + catalog coverage

Open page

why this page is careful

Reference band, not fake proof

  • Best for: Desktop and workstation owners who want a believable starting point for large local reasoning and coding models.
  • Tradeoff: This band is where 70B-class local AI becomes realistic, but even here the benchmark still matters before frontier-scale downloads.
  • Calibration note: High-end desktop class hardware with room for large quantized models.
  • Public-proof boundary: Specific device pages stay gated until shared benchmark evidence is strong enough to index safely.

evidence sources

Evidence sources

  • Benchmark methodology: How the benchmark turns the workstation desktop band into a machine-specific answer. Open page
  • Model provenance review: Why only reviewed catalog entries are used to populate the starter-model guidance. Open page
  • Llama 3.3 70B model page: 70B class start • 40.0 GB minimum Open page
  • gpt-oss-20b model page: 34B class start • 15.5 GB minimum Open page
  • OLMo 3.1 Instruct 32B model page: 34B class start • 19.5 GB minimum Open page