RTX 3090 Ti

Model configurations

Each card is a model and hardware setup, including any documented host RAM, storage or CPU requirements. Select a recipe to update its speeds, context and exact command.

Search configurations

Sort byModelHardware setupTotal memory (GB)RecipeContextDecode (tok/s)Prefill (tok/s)Evidence↑ A–Z

5 configurations · 10 recipes

Gemma 4 12B

RTX 3090 Ti24 GB1 card per machine · 1 machine

TabbyAPIEXL3 4 bpw

70.4decode tok/s

—prefill tok/s

128Kcontext

Earlier check · Sep 2, 2026

CommandFull recipe ›

Gemma 4 26B A4B

RTX 3090 Ti24 GB1 card per machine · 1 machine

TabbyAPIEXL3 5.1 bpw

101.7decode tok/s

2,046prefill tok/s

64Kcontext

Tested on this card · Sep 26, 2026

CommandFull recipe ›

Qwen3.6-35B-A3B

RTX 3090 Ti24 GB1 card per machine · 1 machine

SGLangEXL3 3 bpw

293.1decode tok/s

1,748prefill tok/s

256Kcontext

Tested on this card · Sep 27, 2026

CommandFull recipe ›

Qwen3.8-27B

RTX 3090 Ti24 GB1 card per machine · 1 machine

Exact recipesglang · EXL3 4 bpw · 160K contextsglang · EXL3 3 bpw · 256K contexttabbyapi · EXL3 4 bpw · 128K contexttabbyapi · EXL3 4 bpw · 256K contextvllm · 40K context

130.2decode tok/s

776prefill tok/s

160Kcontext

Tested on this card · Sep 28, 2026

CommandFull recipe ›

Qwen3.8-Flash-Next

RTX 3090 Ti24 GB1 card per machine · 1 machine

Host RAM

≥ 55 GB free

Disk

≥ 86 GB NVMe

Exact recipesglang · EXL3 3.05 bpw · 128K contextsglang · EXL3 3.05 bpw · 256K context

43.4decode tok/s

1,671prefill tok/s

128Kcontext

Tested on this card · Oct 16, 2026

CommandFull recipe ›

Speeds are the selected recipe’s recorded measurements, not estimates; “—” means unmeasured. Evidence distinguishes checks on this card, sibling-card checks, earlier checks and publisher reports.

Sources: NVIDIA, NVIDIA (2), newegg, NVIDIA (3)