Twoody
Local LLM calculator
A local LLM on the RTX 4090
With 24 GB of its own memory: which open-weight models fit the RTX 4090, how fast they write and read, and how it compares with RTX 5090, RTX 3090, RTX 4080 and Radeon RX 7900 XTX. Computed with the memory Twoody's app shows and public llama.cpp benchmarks.
The best local LLM for the RTX 4090
| Memory | Strongest that runs comfortably | Writes | Suggested by Twoody by default |
|---|---|---|---|
| 24 GB | Qwen3.6 35B-A3B One click in Twoody | 100+ tokens/s | Qwen3.5 9B (7.9 GB) |
Speeds are estimates, not measurements made with Twoody, computed from public llama.cpp benchmarks. Figures reviewed on October 16, 2026.
“Strongest that runs comfortably”: the strongest model that fits comfortably in the card's memory and writes at least at your reading speed (5 tokens per second). “Suggested by Twoody by default”: the model the app installs first on a PC with this card and 32 GB of memory, and the memory it needs there.
Twoody is in private beta: leave your email to request access.
Every model that runs well on the RTX 4090
56 of the 62 models the calculator knows (Twoody's catalogue and a few larger ones) run well here, the strongest first: the memory Twoody's app says each one needs on this computer, how fast it writes, and how long it takes to read a document.
| Model | Memory | Writes | Reads a document |
|---|---|---|---|
| Qwen3.6 35B-A3B One click in Twoody | Needs 16.8 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| Qwen3.6 35B-A3B (4 bits) One click in Twoody | Needs 21.2 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| Qwen3.5 35B-A3B One click in Twoody | Needs 22.9 GB of memory Runs — close other apps | 100+ tokens/s | 20 pages in 1 s |
| Qwen3 32B One click in Twoody | Needs 22.7 GB of memory Runs — close other apps | 32–46 tokens/s | 20 pages in 5 s |
| GLM-4 32B One click in Twoody | Needs 19.6 GB of memory Runs comfortably | 32–47 tokens/s | 20 pages in 5 s |
| Nemotron 3 Nano 30B-A3B One click in Twoody | Needs 21.4 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| Gemma 4 31B One click in Twoody | Needs 19.4 GB of memory Runs comfortably | 35–51 tokens/s | 20 pages in 5 s |
| Qwen3 30B-A3B Instruct 2507 One click in Twoody | Needs 19.2 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| Qwen3.5 27B One click in Twoody | Needs 20.5 GB of memory Runs comfortably | 27–54 tokens/s | 20 pages in 5 s |
| Qwen3.8 27B One click in Twoody | Needs 21.4 GB of memory Runs comfortably | 26–52 tokens/s | 20 pages in 5 s |
| Qwen3.6 27B One click in Twoody | Needs 21.5 GB of memory Runs comfortably | 26–51 tokens/s | 20 pages in 5 s |
| Gemma 4 26B (A4B) One click in Twoody | Needs 15.5 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| Mistral Small 3.2 24B One click in Twoody | Needs 17.3 GB of memory Runs comfortably | 45–64 tokens/s | 20 pages in 4 s |
| Qwen3 14B One click in Twoody | Needs 11.2 GB of memory Runs comfortably | 68–98 tokens/s | 20 pages in 2 s |
| Phi-4 14B One click in Twoody | Needs 11.9 GB of memory Runs comfortably | 68–97 tokens/s | 20 pages in 2 s |
| Ministral 3 14B One click in Twoody | Needs 11.6 GB of memory Runs comfortably | 74–106 tokens/s | 20 pages in 2 s |
| Gemma 4 12B One click in Twoody | Needs 7.6 GB of memory Runs comfortably | 83–120 tokens/s | 20 pages in 2 s |
| GLM-4.6V Flash One click in Twoody | Needs 7.9 GB of memory Runs comfortably | 97–139 tokens/s | 20 pages in 2 s |
| Qwen3.5 9B One click in Twoody | Needs 7.9 GB of memory Runs comfortably | 81–163 tokens/s | 20 pages in 2 s |
| GLM-4 9B One click in Twoody | Needs 6.7 GB of memory Runs comfortably | 97–139 tokens/s | 20 pages in 2 s |
| Ministral 3 8B One click in Twoody | Needs 8.4 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| Granite 4.1 8B One click in Twoody | Needs 8 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| Qwen3 8B One click in Twoody | Needs 7.2 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| Llama 3.1 8B One click in Twoody | Needs 6.9 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| Olmo 3 7B One click in Twoody | Needs 8.2 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| Gemma 4 E4B One click in Twoody | Needs 6.6 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| Qwen3.5 4B One click in Twoody | Needs 4.9 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| Qwen3 4B Instruct 2507 One click in Twoody | Needs 4.9 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| Nemotron 3 Nano 4B One click in Twoody | Needs 3.6 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| Ministral 3 3B One click in Twoody | Needs 5 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| Phi-4 mini One click in Twoody | Needs 4.6 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| Llama 3.2 3B One click in Twoody | Needs 3.9 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| Gemma 4 E2B One click in Twoody | Needs 4.7 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| Qwen3.5 2B One click in Twoody | Needs 2.9 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| Qwen3.5 0.8B One click in Twoody | Needs 1.7 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| DeepSeek-R1 Distill Qwen 32B One click in Twoody | Needs 22.8 GB of memory Runs — close other apps | 32–46 tokens/s | 20 pages in 5 s |
| GLM-4.7 Flash One click in Twoody | Needs 18.1 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| Qwen3-Coder 30B-A3B One click in Twoody | Needs 19.1 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| Qwen3 30B-A3B Thinking 2507 One click in Twoody | Needs 19.2 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| Magistral Small 1.2 One click in Twoody | Needs 16.1 GB of memory Runs comfortably | 45–64 tokens/s | 20 pages in 4 s |
| Devstral Small 2 24B One click in Twoody | Needs 16.1 GB of memory Runs comfortably | 45–64 tokens/s | 20 pages in 4 s |
| gpt-oss 20B One click in Twoody | Needs 12 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| DeepSeek-Coder-V2 Lite One click in Twoody | Needs 10.4 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| DeepSeek-R1 Distill Qwen 14B One click in Twoody | Needs 11.7 GB of memory Runs comfortably | 67–96 tokens/s | 20 pages in 2 s |
| Phi-4 reasoning plus One click in Twoody | Needs 11.9 GB of memory Runs comfortably | 68–97 tokens/s | 20 pages in 2 s |
| Ministral 3 14B Reasoning One click in Twoody | Needs 10.5 GB of memory Runs comfortably | 74–106 tokens/s | 20 pages in 2 s |
| GLM-Z1 9B One click in Twoody | Needs 6.7 GB of memory Runs comfortably | 97–139 tokens/s | 20 pages in 2 s |
| DeepSeek-R1 8B One click in Twoody | Needs 7.2 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| DeepSeek-R1 Distill Llama 8B One click in Twoody | Needs 6.9 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| DeepSeek-R1 Distill Qwen 7B One click in Twoody | Needs 5.5 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| Qwen3 4B Thinking 2507 One click in Twoody | Needs 4.9 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| Phi-4 mini reasoning One click in Twoody | Needs 4.6 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| SmolLM3 3B One click in Twoody | Needs 3.2 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| LFM2.5 2.6B One click in Twoody | Needs 2.4 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| MiniCPM5 2B One click in Twoody | Needs 2.5 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
| Llama 3.2 1B One click in Twoody | Needs 1.6 GB of memory Runs comfortably | 100+ tokens/s | 20 pages in 1 s |
Besides these, 2 models would run partly on the processor, much slower, and 4 are too big for this computer. See them all in the calculator.
Each model's memory is what Twoody shows in “Models” for this computer: the same formula, the same files (the one for reading images included), the same conversation length — catalogue checked on October 15, 2026.
RTX 4090 compared with RTX 5090, RTX 3090, RTX 4080 and Radeon RX 7900 XTX
| Computer | Memory | Qwen3.5 9B writes | Reads a document | Strongest that runs comfortably |
|---|---|---|---|---|
| RTX 4090 | 24 GB | 81–163 tokens/s | 20 pages in 2 s | Qwen3.6 35B-A3B 100+ tokens/s |
| RTX 5090 | 32 GB | 100+ tokens/s | 20 pages in 2 s | Qwen3.6 35B-A3B 100+ tokens/s |
| RTX 3090 | 24 GB | 71–143 tokens/s | 20 pages in 4 s | Qwen3.6 35B-A3B 100+ tokens/s |
| RTX 4080 | 16 GB | 60–121 tokens/s | 20 pages in 3 s | Qwen3 14B 50–72 tokens/s |
| Radeon RX 7900 XTX | 24 GB | 78–156 tokens/s | 20 pages in 6 s | Qwen3.6 35B-A3B 100+ tokens/s |
The card's own memory decides which models fit; its bandwidth, how fast they write. Each PC is assumed to have 32 GB of memory beside its card.
This card has 24 GB of memory of its own (VRAM), read at 1,008 GB/s: a model that fits it entirely writes fast; a larger one spills into the computer's memory and slows down a lot.
The tables assume 32 GB of memory in the PC beside the card. The speeds are those of a model running on the card, estimated from public llama.cpp benchmarks.
Twoody for Windows and Linux is in private beta: on Windows 10 and 11 (x64), it runs the model on the RTX 4090 through Vulkan when the driver supports it, and on the processor otherwise; on Linux, on the processor. The speeds on this page are estimates for a model running on the card, not measurements made with Twoody.
Questions about running an LLM locally
Which local LLM for the RTX 4090?
The strongest model that runs comfortably there is Qwen3.6 35B-A3B, which writes 100+ tokens/s. It installs in one click from Twoody's catalogue. On a PC with 32 GB of memory, Twoody suggests Qwen3.5 9B by default, which needs 7.9 GB there: one of the models it has measured with its memory, its tools and your documents.
What fits in the RTX 4090's 24 GB?
56 of the 62 models the calculator knows (Twoody's catalogue and a few larger ones) run well there, on the card; the largest is Qwen3.5 35B-A3B, which needs 22.9 GB. Larger ones spill into the computer's memory and slow down a lot.
How many tokens per second does the RTX 4090 write?
With Qwen3.5 9B, the model Twoody suggests there, it writes 81–163 tokens/s and reads a document of about 20 pages in 2 s. These speeds are estimates from public llama.cpp benchmarks, for a model running on the card, not measurements made with Twoody.
How does the RTX 4090 compare with the RTX 5090, RTX 3090, RTX 4080 and Radeon RX 7900 XTX?
With Qwen3.5 9B, the model Twoody suggests: RTX 4090, 81–163 tokens/s; RTX 5090, 100+ tokens/s; RTX 3090, 71–143 tokens/s; RTX 4080, 60–121 tokens/s; Radeon RX 7900 XTX, 78–156 tokens/s. The card's memory decides which models fit; its bandwidth, how fast they write.
Does Twoody use the RTX 4090?
In the private beta, on Windows 10 and 11 (x64), yes: Twoody runs the model on the card through Vulkan when the driver supports it, and on the processor otherwise. On Linux, it runs the model on the processor. Leave your email to join the waitlist.
Try Twoody before it opens to everyone.
The private beta is free, on Mac, Windows and Linux, and on iPhone and Android with Twoody on your computer. Leave your email to ask for an invitation: we will write to you when it is ready for you.