Which local LLMs can the Radeon RX 7900 XTX run?
Assumes 32 GB of DDR5-5600 system RAM, Windows with this GPU driving the display, 8k context and an f16 KV cache.
Specs
- VRAM
- 24 GB
- Memory bandwidth
- 960.0 GB/s
- FP16 compute
- 122.8 TFLOPS
- Launch year
- 2022
- Price
- $999 launch MSRP
Verdicts at a glance
At Q4_K_M (or the closest available quant) with 8k context.
- Runs great24 modelsRuns well0 modelsRuns slowly0 modelsWon't run5 models2 more models run at a lower quant.
How to read the verdicts
- Runs great
- Fully on the GPU at 20 tok/s or more
- Runs well
- Fully on the GPU at 8–20 tok/s; MoE experts in system RAM at 20 tok/s or more; or at least 90% on the GPU at 8 tok/s or more
- Runs slowly
- 2–8 tok/s; CPU-only; less than 90% on the GPU; or MoE experts in system RAM below 20 tok/s
- Won't run
- Does not fit, or under 2 tok/s
Every model on the Radeon RX 7900 XTX
Dense models split between GPU and CPU slow down sharply — the CPU side sets the pace. MoE models that keep only their experts in system RAM degrade far more gently.
Scroll sideways to see every column.
| Model | Verdict | Speed | Quant | Memory | Runs as | Context | Notes |
|---|---|---|---|---|---|---|---|
| EXAONE 4.0 1.2B | Runs great | est. 570.2 tok/s501.7–638.6calibrated estimate ±12% | Q4_K_M | 2.5 / 24.0 GB | Full GPU | up to 64k | — |
| HyperCLOVA X SEED 1.5B | Runs great | est. 412.5 tok/s363.0–462.0calibrated estimate ±12% | Q4_K_M | 3.0 / 24.0 GB | Full GPU | up to 16k | — |
| Qwen3.5 35B-A3B | Runs great | est. 233.4 tok/s186.7–280.1calibrated estimate ±20% | Q4_K_M | 24.0 / 24.0 GB | Full GPU | up to 8k | — |
| gpt-oss-20b | Runs great | est. 209.9 tok/s167.9–251.8calibrated estimate ±20% | MXFP4 | 13.5 / 24.0 GB | Full GPU | up to 128k |
|
| Qwen3 30B-A3B (2507) | Runs great | est. 170.6 tok/s136.5–204.7calibrated estimate ±20% | Q4_K_M | 20.5 / 24.0 GB | Full GPU | up to 32k | — |
| Gemma 4 26B-A4B | Runs great | est. 166.1 tok/s132.9–199.3calibrated estimate ±20% | Q4_K_M | 18.5 / 24.0 GB | Full GPU | up to 64k | — |
| Kanana 1.5 15.7B-A3B | Runs great | est. 156.4 tok/s125.1–187.7calibrated estimate ±20% | Q4_K_M | 12.7 / 24.0 GB | Full GPU | up to 32k | — |
| DeepSeek R1 Distill Llama 8B | Runs great | est. 124.9 tok/s109.9–139.9calibrated estimate ±12% | Q4_K_M | 7.2 / 24.0 GB | Full GPU | up to 64k |
|
| Kanana 1.5 8B | Runs great | est. 124.9 tok/s109.9–139.9calibrated estimate ±12% | Q4_K_M | 7.2 / 24.0 GB | Full GPU | up to 32k | — |
| Llama 3.1 8B | Runs great | est. 124.9 tok/s109.9–139.9calibrated estimate ±12% | Q4_K_M | 7.2 / 24.0 GB | Full GPU | up to 64k | — |
| Qwen3.5 9B | Runs great | est. 122.0 tok/s107.3–136.6calibrated estimate ±12% | Q4_K_M | 7.3 / 24.0 GB | Full GPU | up to 256k | — |
| Qwen3 8B | Runs great | est. 120.0 tok/s105.6–134.4calibrated estimate ±12% | Q4_K_M | 7.4 / 24.0 GB | Full GPU | up to 32k | — |
| Gemma 4 12B | Runs great | est. 97.8 tok/s86.1–109.5calibrated estimate ±12% | Q4_K_M | 8.8 / 24.0 GB | Full GPU | up to 128k | — |
| Gemma 3 12B | Runs great | est. 95.5 tok/s84.1–107.0calibrated estimate ±12% | Q4_K_M | 9.0 / 24.0 GB | Full GPU | up to 128k | — |
| HyperCLOVA X SEED Think 14B | Runs great | est. 73.4 tok/s51.4–95.5theoretical estimate ±30% | Q4_K_M | 11.4 / 24.0 GB | Full GPU | up to 64k |
|
| Qwen3 14B | Runs great | est. 72.4 tok/s63.7–81.1calibrated estimate ±12% | Q4_K_M | 11.5 / 24.0 GB | Full GPU | up to 32k | — |
| Phi-4 | Runs great | est. 69.8 tok/s61.4–78.2calibrated estimate ±12% | Q4_K_M | 11.9 / 24.0 GB | Full GPU | up to 16k |
|
| Mistral Small 3.2 24B | Runs great | est. 47.8 tok/s42.0–53.5calibrated estimate ±12% | Q4_K_M | 16.8 / 24.0 GB | Full GPU | up to 32k | — |
| Gemma 3 27B | Runs great | est. 43.5 tok/s38.3–48.7calibrated estimate ±12% | Q4_K_M | 18.4 / 24.0 GB | Full GPU | up to 64k | — |
| Qwen3.5 27B | Runs great | est. 42.4 tok/s37.3–47.5calibrated estimate ±12% | Q4_K_M | 18.8 / 24.0 GB | Full GPU | up to 64k | — |
| EXAONE 4.0 32B | Runs great | est. 37.7 tok/s33.2–42.2calibrated estimate ±12% | Q4_K_M | 21.1 / 24.0 GB | Full GPU | up to 32k |
|
| EXAONE 4.5 33B | Runs great | est. 36.4 tok/s32.0–40.7calibrated estimate ±12% | Q4_K_M | 21.8 / 24.0 GB | Full GPU | up to 32k |
|
| Qwen3 32B | Runs great | est. 34.2 tok/s30.1–38.3calibrated estimate ±12% | Q4_K_M | 23.1 / 24.0 GB | Full GPU | up to 8k | — |
| DeepSeek R1 Distill Qwen 32B | Runs great | est. 34.0 tok/s30.0–38.1calibrated estimate ±12% | Q4_K_M | 23.2 / 24.0 GB | Full GPU | up to 8k |
|
| Qwen3.5 122B-A10B | Won't runTry UD-Q2_K_XL (heavy quality loss): Runs slowly | est. 20.7 tok/s14.5–26.8theoretical estimate ±30% | UD-Q2_K_XL | 24.0 / 24.0 GB + 20.2 GB RAM | Partial offload | up to 128k |
|
| Llama 3.3 70B | Won't runTry IQ4_XS: Runs slowly | est. 2.2 tok/s1.8–2.6calibrated estimate ±20% | IQ4_XS | 24.0 / 24.0 GB + 17.8 GB RAM | Partial offload | up to 8k | — |
| gpt-oss-120b | Won't run | — | MXFP4 | needs 64.3 GB | — | — |
|
| Solar Open 100B | Won't run | — | Q4_K_M | needs 64.4 GB | — | — |
|
| Solar Open 2 250B | Won't run | — | IQ4_XS | needs 137.2 GB | — | — |
|
Reasoning models spend extra tokens thinking, so their speed thresholds are 1.5× stricter (30 / 12 / 3 tok/s).
Measured results on the Radeon RX 7900 XTX
Public benchmarks we calibrate against. Their conditions (context, backend, flags) can differ from the estimates above.
| Model | Quant | Backend | Context | Prompt (tok/s) | Generation (tok/s) | Flags | Source | Measured |
|---|---|---|---|---|---|---|---|---|
| llama-2-7b | Q4_0 | llama.cpp | 512 | 3,532.0 | 191.3 | Vulkan | github.com | 2024-12-15 |
Frequently asked questions
What is the largest model that runs entirely on the Radeon RX 7900 XTX?
Qwen3.5 35B-A3B at Q4_K_M (a 22.63 GB file) fits entirely in 24 GB with 8k context, at est. 233.4 tok/s (186.7–280.1, calibrated estimate ±20%).
How many local LLMs run well on the Radeon RX 7900 XTX?
At Q4_K_M with 8k context, out of 29 tracked models: 24 run great, 0 run well, 0 run slowly and 5 won't run.
Can the Radeon RX 7900 XTX run a 70B model like Llama 3.3 70B?
Runs slowly — Partial offload, IQ4_XS: est. 2.2 tok/s (1.8–2.6, calibrated estimate ±20%).
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.