Can I run Qwen3 8B on the GeForce RTX 5090?
Runs greatQ4_K_M fits in VRAM with 24.9 GB to spare: Runs great, 234.3 tok/s measured at 512 context (1 run; 8k estimate 201.1 tok/s, 177.0–225.2, calibrated estimate ±12%).
32 GB VRAM · 1,792.0 GB/s · FP16 104.8 TFLOPS. Assumes 32 GB of DDR5-5600 system RAM, Windows with this GPU driving the display, 8k context and an f16 KV cache. The verdict and quant file above stay the same anywhere from 16 to 128 GB of system RAM.
Every quant of Qwen3 8B on the GeForce RTX 5090
| Quant | File size | Verdict | Speed | Memory | Runs as | Notes |
|---|---|---|---|---|---|---|
| Q2_K | 3.28 GB | Runs great | est. 279.5 tok/s246.0–313.0calibrated estimate ±12% | 5.5 / 32.0 GB | Full GPU | — |
| IQ4_XS | 4.58 GB | Runs great | est. 216.7 tok/s190.7–242.7calibrated estimate ±12% | 6.6 / 32.0 GB | Full GPU | — |
| Q4_K_Mbaseline | 5.03 GB | Runs great | 234.3 tok/s8k estimate 201.1 tok/s (177.0–225.2, calibrated estimate ±12%)measured (1 run, 512 context) | 7.1 / 32.0 GB | Full GPU | — |
| Q8_0 | 8.71 GB | Runs great | 136.2 tok/s8k estimate 126.5 tok/s (111.3–141.7, calibrated estimate ±12%)measured (1 run, 512 context) | 10.4 / 32.0 GB | Full GPU | — |
Models with a thinking mode spend extra tokens thinking, so their speed thresholds are 1.5× stricter (30 / 12 / 3 tok/s). With thinking turned off, a model may do one grade better.
Where the memory goes at Q4_K_M
- Weights
- 4.7 GB
- KV cache
- 1.2 GB
- Compute buffer
- 0.6 GB
- OS reserve
- 0.6 GB
- Free
- 24.9 GB
0.36 GB of the 5.03 GB file stays off the graphics card: llama.cpp keeps the input embedding table in system RAM.
Context length vs. KV cache
| Context | KV f16 | KV q8_0 | KV q4_0 |
|---|---|---|---|
| 4k | 5.8 GBRuns great est. 222.6 tok/s195.9–249.4calibrated estimate ±12% | 5.5 GBRuns great est. 234.3 tok/s206.2–262.4calibrated estimate ±12% | 5.4 GBRuns great est. 241.1 tok/s212.2–270.1calibrated estimate ±12% |
| 8k | 6.5 GBRuns great 234.3 tok/s8k estimate 201.1 tok/s (177.0–225.2, calibrated estimate ±12%)measured (1 run, 512 context) | 5.9 GBRuns great est. 221.0 tok/s194.5–247.5calibrated estimate ±12% | 5.6 GBRuns great est. 233.4 tok/s205.4–261.4calibrated estimate ±12% |
| 16k | 7.7 GBRuns great est. 168.5 tok/s148.3–188.7calibrated estimate ±12% | 6.6 GBRuns great est. 198.4 tok/s174.6–222.2calibrated estimate ±12% | 6.0 GBRuns great est. 219.4 tok/s193.0–245.7calibrated estimate ±12% |
| 32k | 10.3 GBRuns great est. 127.2 tok/s111.9–142.5calibrated estimate ±12% | 8.1 GBRuns great est. 164.7 tok/s145.0–184.5calibrated estimate ±12% | 6.8 GBRuns great est. 195.8 tok/s172.3–219.3calibrated estimate ±12% |
S = Runs great · A = Runs well · B = Runs slowly · F = Won't run
At Q4_K_M the verdict stays Runs great up to 32k context with an f16 KV cache, and up to 32k with q8_0.
Estimated speed
- Generation
- 234.3 tok/s measured at 512 context (1 run; 8k estimate 201.1 tok/s, 177.0–225.2, calibrated estimate ±12%)
- Prompt processing
- about 13,385 tok/s (8,031–18,739, rough estimate ±40%)
Measured on this exact combination
Public benchmarks we calibrate against. Their conditions (context, backend, flags) can differ from the estimates above.
| Hardware | Quant | Backend | Context | Prompt (tok/s) | Generation (tok/s) | Flags | Source | Measured |
|---|---|---|---|---|---|---|---|---|
| GeForce RTX 5090 | Q4_K_XL | llama.cpp | 16k | — | 145.3 | — | hardware-corner.net | 2026-08-09 |
| GeForce RTX 5090 | Q4_K_XL | llama.cpp | 4k | 11,933.0 | 200.4 | CUDA | hardware-corner.net | 2026-03-02 |
| GeForce RTX 5090 | Q4_K_M | llama.cpp | 512 | — | 234.3 | CUDA, KV q8_0 | github.com | 2026-10-07 |
| GeForce RTX 5090 | Q8_0 | llama.cpp | 512 | 10,787.0 | 136.2 | CUDA | github.com | 2025-10-16 |
If this is not enough
A larger model that also runs well on the GeForce RTX 5090
Runs wellKolibri 1: Runs well, est. 83.3 tok/s (58.3–108.3, theoretical estimate ±30%)
A larger quant that still runs great
Runs greatQ8_0 (8.71 GB): Runs great, 136.2 tok/s measured at 512 context (1 run; 8k estimate 126.5 tok/s, 111.3–141.7, calibrated estimate ±12%)
How to run it
Commands for the Q4_K_M file. Both tools download from Hugging Face on first run.
File: Qwen3-8B-Q4_K_M.gguf (5.03 GB) from unsloth/Qwen3-8B-GGUF.
llama-server -hf unsloth/Qwen3-8B-GGUF:Q4_K_M -c 8192 -ngl all -fa on- -ngl all loads every layer on the GPU.
- -fa on enables flash attention (needed for KV cache quantization).
- --cache-type-k q8_0 --cache-type-v q8_0 shrinks the KV cache to 0.65 GB at 8k context.
PowerShell (quit the Ollama tray app first):
$env:OLLAMA_CONTEXT_LENGTH="8192"; ollama serve
ollama run hf.co/unsloth/Qwen3-8B-GGUF:Q4_K_M- Ollama defaults to 4k context on GPUs under 24 GB. Set OLLAMA_CONTEXT_LENGTH when starting the server, or /set parameter num_ctx inside the chat.
/set parameter num_ctx 8192
Related pages
Guide: How much VRAM do local LLMs need? GPU requirements for 8–32 GB (2026)
Other models on the GeForce RTX 5090
Qwen3 8B on other hardware
Frequently asked questions
Can I run Qwen3 8B on the GeForce RTX 5090?
Q4_K_M fits in VRAM with 24.9 GB to spare: Runs great, 234.3 tok/s measured at 512 context (1 run; 8k estimate 201.1 tok/s, 177.0–225.2, calibrated estimate ±12%). Q4_K_M needs 6.5 GB at 8k context; this setup has 31.4 GB of usable memory and 28.0 GB of free system RAM.
How much context can Qwen3 8B use on the GeForce RTX 5090?
At Q4_K_M the verdict stays Runs great up to 32k context with an f16 KV cache (4.8 GB of KV) and up to 32k with a q8_0 KV cache (2.6 GB). The model supports up to 32k.
Which quant should I use, and how fast is it?
Q4_K_M (5.03 GB) is the recommended balance: Runs great, 234.3 tok/s measured at 512 context (1 run; 8k estimate 201.1 tok/s, 177.0–225.2, calibrated estimate ±12%). The largest tracked file that still gets Runs great is Q8_0 (8.71 GB), 136.2 tok/s measured at 512 context (1 run; 8k estimate 126.5 tok/s, 111.3–141.7, calibrated estimate ±12%).
Speeds are estimates from memory bandwidth, calibrated against public benchmarks, and each one comes with an error band and a confidence label. Real results vary with drivers, backend, context length and thermals.