RX 9070 XT vs RTX 5070 for local LLMs
Assumes 32 GB of DDR5-5600 system RAM, Windows with this GPU driving the display, 8k context and an f16 KV cache.
What to know before choosing
With near-equal memory bandwidth, the verdicts here split on memory size. CanRun counts 15.4 GB usable on the RX 9070 XT and 11.4 GB on the RTX 5070. gpt-oss-20b needs 12.3 GB at 8k, so it fits fully on the RX 9070 XT and runs great. On the RTX 5070 some experts go to system RAM; it runs well, a theoretical estimate (±30%). Kanana 1.5 15.7B-A3B, a Korean model, splits the same way, again theoretical. Mistral Small 3.2 24B outgrows both: the RX 9070 XT keeps most of it on the GPU and runs it well, while the RTX 5070 spills far more and runs it slowly.
At 8k, Phi-4 runs great on the RTX 5070 with little margin (check ollama ps). At 16k, its maximum, it needs 12.8 GB with an f16 KV cache: the RTX 5070 moves part of it to system RAM while the RX 9070 XT keeps it all on the GPU. A q8_0 KV cache fits it on the RTX 5070 too.
The spec table's FP16 figure is much higher on the RX 9070 XT, yet the speed section's prompt-processing estimates are close. The two makers count FP16 peaks differently; AMD's figure is a dual-issue peak. CanRun's prompt-processing estimate uses per-maker factors fitted to public llama.cpp results and is only a rough guide.
Speed
The RX 9070 XT and the RTX 5070 generate text at about the same speed (median of our estimates over the 27 models that fit entirely in memory on both at Q4_K_M or the closest available quant).
- Reference model
- Qwen3 14B at Q4_K_M, the largest model that fits entirely in memory on both at Q4_K_M or the closest available quant
- RX 9070 XT
- est. 47.4 tok/s (41.7–53.1, calibrated estimate ±12%)
- RTX 5070
- est. 45.5 tok/s (40.0–50.9, calibrated estimate ±12%)
- Prompt processing on Qwen3 14B
- RX 9070 XT: about 2,237 tok/s (1,342–3,131, rough estimate ±40%); RTX 5070: about 2,228 tok/s (1,337–3,119, rough estimate ±40%)
Specs side by side
| Spec | RX 9070 XT | RTX 5070 | Difference |
|---|---|---|---|
| Memory | 16 GB | 12 GB | +4 GB on the RX 9070 XT |
| Memory bandwidth | 645.0 GB/s | 672.0 GB/s | 1.04× on the RTX 5070 |
| FP16 compute | 97.3 TFLOPS | 30.9 TFLOPS | 3.15× on the RX 9070 XT |
| Launch year | 2025 | 2025 | — |
| Price | $689 street (as of 2026-08-09) | $629 street (as of 2026-08-09) | — |
| Public benchmark results | 1 result | 3 results | — |
Verdicts at a glance
Out of 58 models at Q4_K_M (or the closest available quant) with 8k context.
- RX 9070 XTRuns great29 modelsRuns well8 modelsRuns slowly11 modelsWon't run10 models
- RTX 5070Runs great27 modelsRuns well8 modelsRuns slowly13 modelsWon't run10 models
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
Where the verdicts differ
4 models get a different verdict on the two.
Scroll sideways to see every column.
| Model | RX 9070 XT verdict | RX 9070 XT speed | RTX 5070 verdict | RTX 5070 speed |
|---|---|---|---|---|
| gpt-oss-20bMXFP4 | Runs great | est. 139.9 tok/s111.9–167.9calibrated estimate ±20% | Runs wellgpt-oss-20b verdict on the RTX 5070 | est. 91.6 tok/s64.1–119.1theoretical estimate ±30% |
| Kanana 1.5 15.7B-A3BQ4_K_M | Runs great | est. 101.1 tok/s80.9–121.4calibrated estimate ±20% | Runs well | est. 95.9 tok/s67.1–124.7theoretical estimate ±30% |
| K2-Horizon MoVA 36B-A4BQ4_K_M | Runs well | est. 32.7 tok/s22.9–42.5theoretical estimate ±30% | Runs slowly | est. 29.0 tok/s20.3–37.7theoretical estimate ±30% |
| Mistral Small 3.2 24BQ4_K_M | Runs well | est. 19.7 tok/s15.8–23.6calibrated estimate ±20% | Runs slowly | est. 7.1 tok/s5.7–8.5calibrated estimate ±20% |
Every model on both
Scroll sideways to see every column.
| Model | RX 9070 XT verdict | RX 9070 XT speed | RTX 5070 verdict | RTX 5070 speed |
|---|---|---|---|---|
| gpt-oss-20bMXFP4 | Runs great | est. 139.9 tok/s111.9–167.9calibrated estimate ±20% | Runs wellgpt-oss-20b verdict on the RTX 5070 | est. 91.6 tok/s64.1–119.1theoretical estimate ±30% |
| Kanana 1.5 15.7B-A3BQ4_K_M | Runs great | est. 101.1 tok/s80.9–121.4calibrated estimate ±20% | Runs well | est. 95.9 tok/s67.1–124.7theoretical estimate ±30% |
| Qwen3 14BQ4_K_M | Runs great | est. 47.4 tok/s41.7–53.1calibrated estimate ±12% | Runs greatQwen3 14B verdict on the RTX 5070 | est. 45.5 tok/s40.0–50.9calibrated estimate ±12% |
| DeepSeek R1 Distill Qwen 14BQ4_K_M | Runs great | est. 46.2 tok/s40.7–51.8calibrated estimate ±12% | Runs great | est. 44.4 tok/s39.0–49.7calibrated estimate ±12% |
| HyperCLOVA X SEED Think 14BQ4_K_M | Runs great | est. 48.1 tok/s33.7–62.5theoretical estimate ±30% | Runs great | est. 46.1 tok/s32.3–60.0theoretical estimate ±30% |
| Phi-4Q4_K_M | Runs great | est. 45.7 tok/s40.2–51.2calibrated estimate ±12% | Runs great | est. 43.8 tok/s38.6–49.1calibrated estimate ±12% |
| Gemma 3 12BQ4_K_M | Runs great | est. 62.6 tok/s55.0–70.1calibrated estimate ±12% | Runs greatGemma 3 12B verdict on the RTX 5070 | est. 60.0 tok/s52.8–67.2calibrated estimate ±12% |
| Mellum2.1 12B-A2.5B ThinkingQ4_K_M | Runs great | est. 149.3 tok/s119.4–179.1calibrated estimate ±20% | Runs great | est. 196.5 tok/s157.2–235.8calibrated estimate ±20% |
| Gemma 4 12BQ4_K_M | Runs great | est. 67.6 tok/s59.5–75.7calibrated estimate ±12% | Runs great | est. 64.8 tok/s57.1–72.6calibrated estimate ±12% |
| Qwen3.5 9BQ4_K_M | Runs great | est. 79.9 tok/s70.3–89.4calibrated estimate ±12% | Runs great | est. 76.6 tok/s67.4–85.8calibrated estimate ±12% |
| K2-Horizon 7BQ4_K_M | Runs great | est. 72.1 tok/s63.5–80.8calibrated estimate ±12% | Runs great | est. 69.2 tok/s60.9–77.5calibrated estimate ±12% |
| Qwen3 8BQ4_K_M | Runs great | est. 78.6 tok/s69.2–88.0calibrated estimate ±12% | Runs great | est. 75.4 tok/s66.4–84.5calibrated estimate ±12% |
| DeepSeek R1 0528 Qwen3 8BQ4_K_M | Runs great | est. 78.6 tok/s69.2–88.0calibrated estimate ±12% | Runs great | est. 75.4 tok/s66.4–84.5calibrated estimate ±12% |
| DeepSeek R1 Distill Llama 8BQ4_K_M | Runs great | est. 81.8 tok/s72.0–91.6calibrated estimate ±12% | Runs great | est. 78.5 tok/s69.1–87.9calibrated estimate ±12% |
| Kanana 1.5 8BQ4_K_M | Runs great | est. 81.8 tok/s72.0–91.6calibrated estimate ±12% | Runs great | est. 78.5 tok/s69.1–87.9calibrated estimate ±12% |
| Llama 3.1 8BQ4_K_M | Runs great | est. 81.8 tok/s72.0–91.6calibrated estimate ±12% | Runs greatLlama 3.1 8B verdict on the RTX 5070 | est. 78.5 tok/s69.1–87.9calibrated estimate ±12% |
| Gemma 4 E4BQ4_K_M | Runs great | est. 121.6 tok/s85.1–158.0theoretical estimate ±30% | Runs great | est. 123.5 tok/s86.4–160.5theoretical estimate ±30% |
| Ling-3.0 TinyQ4_K_M | Runs great | est. 222.2 tok/s155.5–288.8theoretical estimate ±30% | Runs great | est. 318.9 tok/s223.2–414.6theoretical estimate ±30% |
| Gemma 4 E2BQ4_K_M | Runs great | est. 203.6 tok/s142.5–264.6theoretical estimate ±30% | Runs great | est. 215.2 tok/s150.7–279.8theoretical estimate ±30% |
| Qwen3.5 4BQ4_K_M | Runs great | est. 158.2 tok/s139.2–177.2calibrated estimate ±12% | Runs great | est. 151.8 tok/s133.6–170.0calibrated estimate ±12% |
| Nanbeige4.2 3BQ4_K_M | Runs great | est. 80.0 tok/s56.0–104.0theoretical estimate ±30% | Runs great | est. 76.8 tok/s53.7–99.8theoretical estimate ±30% |
| Spark-X2.5 4BQ4_K_M | Runs great | est. 168.9 tok/s148.6–189.2calibrated estimate ±12% | Runs great | est. 162.1 tok/s142.6–181.5calibrated estimate ±12% |
| Llama 3.2 3BQ4_K_M | Runs great | est. 165.6 tok/s145.8–185.5calibrated estimate ±12% | Runs great | est. 158.9 tok/s139.9–178.0calibrated estimate ±12% |
| LFM2.5 2.6BQ4_K_M | Runs great | est. 271.7 tok/s190.2–353.2theoretical estimate ±30% | Runs great | est. 260.7 tok/s182.5–338.9theoretical estimate ±30% |
| MiniCPM5 2BQ4_K_M | Runs great | est. 256.3 tok/s225.6–287.1calibrated estimate ±12% | Runs great | est. 246.0 tok/s216.5–275.5calibrated estimate ±12% |
| HyperCLOVA X SEED 1.5BQ4_K_M | Runs great | est. 270.0 tok/s237.6–302.4calibrated estimate ±12% | Runs great | est. 259.1 tok/s228.0–290.2calibrated estimate ±12% |
| EXAONE 4.0 1.2BQ4_K_M | Runs great | est. 373.3 tok/s328.5–418.0calibrated estimate ±12% | Runs great | est. 358.2 tok/s315.2–401.2calibrated estimate ±12% |
| Llama 3.2 1BQ4_K_M | Runs great | est. 454.5 tok/s400.0–509.1calibrated estimate ±12% | Runs great | est. 436.2 tok/s383.8–488.5calibrated estimate ±12% |
| MiniCPM5 1BQ4_K_M | Runs great | est. 550.0 tok/s484.0–616.0calibrated estimate ±12% | Runs great | est. 527.8 tok/s464.4–591.1calibrated estimate ±12% |
| K2-Horizon MoVA 36B-A4BQ4_K_M | Runs well | est. 32.7 tok/s22.9–42.5theoretical estimate ±30% | Runs slowly | est. 29.0 tok/s20.3–37.7theoretical estimate ±30% |
| Qwen3.5 35B-A3BQ4_K_M | Runs well | est. 64.1 tok/s44.9–83.3theoretical estimate ±30% | Runs well | est. 63.3 tok/s44.3–82.3theoretical estimate ±30% |
| Qwen3.6 35B-A3BQ4_K_M | Runs well | est. 69.3 tok/s48.5–90.1theoretical estimate ±30% | Runs well | est. 68.6 tok/s48.0–89.2theoretical estimate ±30% |
| MiniCPM-V 4.7 35B-A3BQ4_K_M | Runs well | est. 73.2 tok/s51.2–95.1theoretical estimate ±30% | Runs well | est. 72.5 tok/s50.7–94.2theoretical estimate ±30% |
| Nemotron 3.5 Lightning 30B-A3BQ4_K_M | Runs well | est. 55.0 tok/s38.5–71.5theoretical estimate ±30% | Runs well | est. 47.5 tok/s33.3–61.8theoretical estimate ±30% |
| Qwen3 30B-A3B (2507)Q4_K_M | Runs well | est. 57.0 tok/s39.9–74.2theoretical estimate ±30% | est. 50.5 tok/s35.3–65.6theoretical estimate ±30% | |
| Gemma 4 26B-A4BQ4_K_M | Runs well | est. 67.6 tok/s47.3–87.9theoretical estimate ±30% | Runs well | est. 53.6 tok/s37.5–69.7theoretical estimate ±30% |
| Mistral Small 3.2 24BQ4_K_M | Runs well | est. 19.7 tok/s15.8–23.6calibrated estimate ±20% | Runs slowly | est. 7.1 tok/s5.7–8.5calibrated estimate ±20% |
| EXAONE 4.5 33BQ4_K_M | Runs slowly | est. 6.2 tok/s4.9–7.4calibrated estimate ±20% | Runs slowly | est. 4.1 tok/s3.2–4.9calibrated estimate ±20% |
| DeepSeek R1 Distill Qwen 32BQ4_K_M | Runs slowly | est. 4.9 tok/s3.9–5.8calibrated estimate ±20% | Runs slowly | est. 3.4 tok/s2.7–4.0calibrated estimate ±20% |
| Qwen3 32BQ4_K_M | Runs slowly | est. 4.9 tok/s3.9–5.9calibrated estimate ±20% | Runs slowly | est. 3.4 tok/s2.7–4.1calibrated estimate ±20% |
| EXAONE 4.0 32BQ4_K_M | Runs slowly | est. 6.9 tok/s5.5–8.2calibrated estimate ±20% | Runs slowly | est. 4.3 tok/s3.5–5.2calibrated estimate ±20% |
| Gemma 4 31BQ4_K_M | Runs slowly | est. 7.9 tok/s6.3–9.5calibrated estimate ±20% | Runs slowly | est. 4.7 tok/s3.8–5.7calibrated estimate ±20% |
| Muse Glimmer 30BQ4_K_M | Runs slowly | est. 11.3 tok/s9.0–13.6calibrated estimate ±20% | Runs slowly | est. 5.8 tok/s4.7–7.0calibrated estimate ±20% |
| Granite 4.2 30BQ4_K_M | Runs slowly | est. 6.4 tok/s5.2–7.7calibrated estimate ±20% | Runs slowly | est. 4.0 tok/s3.2–4.9calibrated estimate ±20% |
| Qwen3.5 27BQ4_K_M | Runs slowly | est. 10.5 tok/s8.4–12.6calibrated estimate ±20% | Runs slowly | est. 5.6 tok/s4.5–6.7calibrated estimate ±20% |
| Qwen3.6 27BQ4_K_M | Runs slowly | est. 11.3 tok/s9.1–13.6calibrated estimate ±20% | Runs slowly | est. 5.8 tok/s4.6–6.9calibrated estimate ±20% |
| Qwen3.8 27BQ4_K_M | Runs slowly | est. 9.7 tok/s7.8–11.7calibrated estimate ±20% | Runs slowly | est. 5.3 tok/s4.3–6.4calibrated estimate ±20% |
| Gemma 3 27BQ4_K_M | Runs slowly | est. 11.7 tok/s9.3–14.0calibrated estimate ±20% | Runs slowly | est. 5.8 tok/s4.7–7.0calibrated estimate ±20% |
| GLM-5.3 FlashQ4_K_M | Won't run | — | Won't run | — |
| DeepSeek V4 Flash 0731MXFP4 | Won't run | — | Won't run | — |
| Solar Open 2 250BIQ4_XS | Won't run | — | Won't run | — |
| Ling-3.0 FlashQ4_K_M | Won't run | — | Won't run | — |
| Qwen3.5 122B-A10BQ4_K_M | Won't run | — | Won't run | — |
| Laguna S 2.1Q4_K_M | Won't run | — | Won't run | — |
| gpt-oss-120bMXFP4 | Won't run | — | Won't rungpt-oss-120b verdict on the RTX 5070 | — |
| Solar Open 100BQ4_K_M | Won't runTry Q2_K (heavy quality loss): Runs slowly | est. 19.6 tok/s13.7–25.5theoretical estimate ±30% | Won't run | — |
| Kolibri 1Q4_K_M | Won't runTry Q2_K (heavy quality loss): Runs well | est. 65.5 tok/s45.8–85.1theoretical estimate ±30% | Won't runTry Q2_K (heavy quality loss): Runs well | est. 69.7 tok/s48.8–90.6theoretical estimate ±30% |
| Llama 3.3 70BQ4_K_M | Won't runTry Q2_K (heavy quality loss): Runs slowly | est. 2.6 tok/s2.1–3.2calibrated estimate ±20% | Won't runTry Q2_K (heavy quality loss): Runs slowly | est. 2.1 tok/s1.7–2.6calibrated estimate ±20% |
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.
Public benchmarks
Measured results on the RX 9070 XT
| Model | Quant | Backend | Context | Prompt (tok/s) | Generation (tok/s) | Flags | Source | Measured |
|---|---|---|---|---|---|---|---|---|
| llama-2-7b | Q4_0 | llama.cpp | 512 | 5,048.0 | 131.5 | Vulkan | github.com | 2026-01-24 |
Measured results on the RTX 5070
| Model | Quant | Backend | Context | Prompt (tok/s) | Generation (tok/s) | Flags | Source | Measured |
|---|---|---|---|---|---|---|---|---|
| llama-2-7b | Q4_0 | llama.cpp | 512 | 5,185.0 | 127.5 | — | github.com | 2025-08-20 |
| Qwen3.5 9B | Q8_0 | llama.cpp | 32k | 3,001.0 | 67.3 | — | localmaxxing.com | 2026-09-10 |
| Qwen3 8B | Q4_K_XL | llama.cpp | 4k | 3,488.0 | 85.8 | CUDA | hardware-corner.net | 2025-12-09 |
Public benchmarks we calibrate against. Their conditions (context, backend, flags) can differ from the estimates above.
Frequently asked questions
Which is faster for local LLMs, the RX 9070 XT or the RTX 5070?
The RX 9070 XT and the RTX 5070 generate text at about the same speed (median of our estimates over the 27 models that fit entirely in memory on both at Q4_K_M or the closest available quant). Estimated speed on Qwen3 14B at Q4_K_M: the RX 9070 XT est. 47.4 tok/s (41.7–53.1, calibrated estimate ±12%); the RTX 5070 est. 45.5 tok/s (40.0–50.9, calibrated estimate ±12%).
What is the largest model each can run entirely in memory?
The RX 9070 XT fits Muse Glimmer 30B at IQ4_XS (15.44 GB file) entirely in 16 GB, at est. 31.5 tok/s (27.7–35.3, calibrated estimate ±12%). The RTX 5070 fits Kanana 1.5 15.7B-A3B at IQ4_XS (8.71 GB file) entirely in 12 GB, at est. 130.8 tok/s (104.6–157.0, calibrated estimate ±20%).
How many local LLMs run well on each?
At Q4_K_M with 8k context, out of 58 tracked models: the RX 9070 XT — 29 Runs great, 8 Runs well, 11 Runs slowly, 10 Won't run; the RTX 5070 — 27 Runs great, 8 Runs well, 13 Runs slowly, 10 Won't run. 4 models get a different verdict.
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.