Which local LLMs can the Radeon RX 9070 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: 16 GB
Memory bandwidth: 645.0 GB/s
FP16 compute: 72.3 TFLOPS
Launch year: 2025
Price: $639 street (as of 2026-08-09)
Also written as: 9070

Verdicts at a glance

At Q4_K_M (or the closest available quant) with 8k context.

Runs great29 modelsRuns well8 modelsRuns slowly12 modelsWon't run10 models3 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 9070

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.

59 models sorted by verdict and speed
Model	Verdict	Speed	Quant	Memory	Runs as	Context	Notes
MiniCPM5 1B	Runs great	est. 550.0 tok/s484.0–616.0calibrated estimate ±12%	Q4_K_M	1.9 / 16.0 GB	Full GPU	up to 128k	thinking mode
Llama 3.2 1B	Runs great	est. 454.5 tok/s400.0–509.1calibrated estimate ±12%	Q4_K_M	2.2 / 16.0 GB	Full GPU	up to 128k	—
EXAONE 4.0 1.2B	Runs great	est. 373.3 tok/s328.5–418.0calibrated estimate ±12%	Q4_K_M	2.5 / 16.0 GB	Full GPU	up to 64k	thinking mode
LFM2.5 2.6B	Runs great	est. 271.7 tok/s190.2–353.2theoretical estimate ±30%	Q4_K_M	3.0 / 16.0 GB	Full GPU	up to 128k	thinking mode
HyperCLOVA X SEED 1.5B	Runs great	est. 270.0 tok/s237.6–302.4calibrated estimate ±12%	Q4_K_M	3.0 / 16.0 GB	Full GPU	up to 16k	—
MiniCPM5 2B	Runs great	est. 256.3 tok/s225.6–287.1calibrated estimate ±12%	Q4_K_M	2.9 / 16.0 GB	Full GPU	up to 128k	thinking mode
Ling-3.0 Tiny	Runs great	est. 222.2 tok/s155.5–288.8theoretical estimate ±30%	Q4_K_M	5.9 / 16.0 GB	Full GPU	up to 128k	thinking mode
Gemma 4 E2B	Runs great	est. 203.6 tok/s142.5–264.6theoretical estimate ±30%	Q4_K_M	2.7 / 16.0 GB	Full GPU	up to 128k	thinking mode
Spark-X2.5 4B	Runs great	est. 168.9 tok/s148.6–189.2calibrated estimate ±12%	Q4_K_M	4.1 / 16.0 GB	Full GPU	up to 256k	thinking mode
Llama 3.2 3B	Runs great	est. 165.6 tok/s145.8–185.5calibrated estimate ±12%	Q4_K_M	4.1 / 16.0 GB	Full GPU	up to 64k	—
Qwen3.5 4B	Runs great	est. 158.2 tok/s139.2–177.2calibrated estimate ±12%	Q4_K_M	4.2 / 16.0 GB	Full GPU	up to 256k	thinking mode
Mellum2.1 12B-A2.5B Thinking	Runs great	est. 149.3 tok/s119.4–179.1calibrated estimate ±20%	Q4_K_M	9.2 / 16.0 GB	Full GPU	up to 128k	thinking mode
gpt-oss-20b	Runs great	est. 139.9 tok/s111.9–167.9calibrated estimate ±20%	MXFP4	12.9 / 16.0 GB	Full GPU	up to 64k	thinking mode
Gemma 4 E4B	Runs great	est. 121.6 tok/s85.1–158.0theoretical estimate ±30%	Q4_K_M	4.3 / 16.0 GB	Full GPU	up to 128k	thinking mode
Kanana 1.5 15.7B-A3B	Runs great	est. 101.1 tok/s80.9–121.4calibrated estimate ±20%	Q4_K_M	12.5 / 16.0 GB	Full GPU	up to 32k	—
Qwen3 8B	Runs great	99.3 tok/s8k estimate 78.6 tok/s (69.2–88.0, calibrated estimate ±12%)measured (2 runs, 512 context)	Q4_K_M	7.1 / 16.0 GB	Full GPU	up to 32k	thinking mode
DeepSeek R1 Distill Llama 8B	Runs great	est. 81.8 tok/s72.0–91.6calibrated estimate ±12%	Q4_K_M	6.9 / 16.0 GB	Full GPU	up to 64k	thinking mode
Kanana 1.5 8B	Runs great	est. 81.8 tok/s72.0–91.6calibrated estimate ±12%	Q4_K_M	6.9 / 16.0 GB	Full GPU	up to 32k	—
Llama 3.1 8B	Runs great	est. 81.8 tok/s72.0–91.6calibrated estimate ±12%	Q4_K_M	6.9 / 16.0 GB	Full GPU	up to 64k	—
Nanbeige4.2 3B	Runs great	est. 80.0 tok/s56.0–104.0theoretical estimate ±30%	Q4_K_M	5.0 / 16.0 GB	Full GPU	up to 32k	thinking mode
Qwen3.5 9B	Runs great	est. 79.9 tok/s70.3–89.4calibrated estimate ±12%	Q4_K_M	6.6 / 16.0 GB	Full GPU	up to 128k	thinking mode
DeepSeek R1 0528 Qwen3 8B	Runs great	est. 78.6 tok/s69.2–88.0calibrated estimate ±12%	Q4_K_M	7.1 / 16.0 GB	Full GPU	up to 32k	thinking mode
K2-Horizon 7B	Runs great	est. 72.1 tok/s63.5–80.8calibrated estimate ±12%	Q4_K_M	7.4 / 16.0 GB	Full GPU	up to 32k	thinking mode
Gemma 4 12B	Runs great	est. 67.6 tok/s59.5–75.7calibrated estimate ±12%	Q4_K_M	8.4 / 16.0 GB	Full GPU	up to 256k	thinking mode
Gemma 3 12B	Runs great	est. 62.6 tok/s55.0–70.1calibrated estimate ±12%	Q4_K_M	9.0 / 16.0 GB	Full GPU	up to 64k	—
HyperCLOVA X SEED Think 14B	Runs great	est. 48.1 tok/s33.7–62.5theoretical estimate ±30%	Q4_K_M	11.0 / 16.0 GB	Full GPU	up to 32k	thinking mode
Qwen3 14B	Runs great	est. 47.4 tok/s41.7–53.1calibrated estimate ±12%	Q4_K_M	11.1 / 16.0 GB	Full GPU	up to 32k	thinking mode
DeepSeek R1 Distill Qwen 14B	Runs great	est. 46.2 tok/s40.7–51.8calibrated estimate ±12%	Q4_K_M	11.3 / 16.0 GB	Full GPU	up to 16k	thinking mode
Phi-4	Runs great	est. 45.7 tok/s40.2–51.2calibrated estimate ±12%	Q4_K_M	11.6 / 16.0 GB	Full GPU	up to 16k	—
MiniCPM-V 4.7 35B-A3B	Runs well	est. 73.2 tok/s51.2–95.1theoretical estimate ±30%	Q4_K_M	15.4 / 16.0 GB + 7.1 GB RAM	MoE experts in RAM	up to 256k	Not fully on the GPU — Runs great needs the whole model in GPU memory thinking mode
Qwen3.6 35B-A3B	Runs well	est. 69.3 tok/s48.5–90.1theoretical estimate ±30%	Q4_K_M	15.1 / 16.0 GB + 8.5 GB RAM	MoE experts in RAM	up to 256k	Not fully on the GPU — Runs great needs the whole model in GPU memory Q2_K (heavy quality loss): Runs great thinking mode
Gemma 4 26B-A4B	Runs well	est. 67.6 tok/s47.3–87.9theoretical estimate ±30%	Q4_K_M	15.4 / 16.0 GB + 2.9 GB RAM	MoE experts in RAM	up to 128k	Not fully on the GPU — Runs great needs the whole model in GPU memory Try IQ4_XS: Runs great thinking mode
Qwen3.5 35B-A3B	Runs well	est. 64.1 tok/s44.9–83.3theoretical estimate ±30%	Q4_K_M	15.1 / 16.0 GB + 8.9 GB RAM	MoE experts in RAM	up to 256k	Not fully on the GPU — Runs great needs the whole model in GPU memory UD-Q2_K_XL (heavy quality loss): Runs great thinking mode
Qwen3 30B-A3B (2507)	Runs well	est. 57.0 tok/s39.9–74.2theoretical estimate ±30%	Q4_K_M	15.6 / 16.0 GB + 4.9 GB RAM	MoE experts in RAM	up to 64k	Not fully on the GPU — Runs great needs the whole model in GPU memory Q2_K (heavy quality loss): Runs great
Nemotron 3.5 Lightning 30B-A3B	Runs well	est. 55.0 tok/s38.5–71.5theoretical estimate ±30%	Q4_K_M	14.8 / 16.0 GB + 11.9 GB RAM	MoE experts in RAM	up to 256k	Not fully on the GPU — Runs great needs the whole model in GPU memory thinking mode
K2-Horizon MoVA 36B-A4B	Runs well	est. 32.7 tok/s22.9–42.5theoretical estimate ±30%	Q4_K_M	15.5 / 16.0 GB + 9.6 GB RAM	MoE experts in RAM	up to 8k	Not fully on the GPU — Runs great needs the whole model in GPU memory thinking mode
Mistral Small 3.2 24B	Runs well	est. 19.7 tok/s15.8–23.6calibrated estimate ±20%	Q4_K_M	16.0 / 16.0 GB + 0.8 GB RAM	Partial offload	up to 8k	Try IQ4_XS: Runs great
Gemma 3 27B	Runs slowly	est. 11.7 tok/s9.3–14.0calibrated estimate ±20%	Q4_K_M	16.0 / 16.0 GB + 2.4 GB RAM	Partial offload	up to 64k	Less than 90% on the GPU — a GPU/CPU split is capped at Runs slowly Try IQ4_XS: Runs well Q2_K (heavy quality loss): Runs great
Qwen3.6 27B	Runs slowly	est. 11.3 tok/s9.1–13.6calibrated estimate ±20%	Q4_K_M	16.0 / 16.0 GB + 2.5 GB RAM	Partial offload	up to 64k	Try IQ4_XS: Runs well UD-Q2_K_XL (heavy quality loss): Runs great thinking mode
Muse Glimmer 30B	Runs slowly	est. 11.3 tok/s9.0–13.6calibrated estimate ±20%	Q4_K_M	16.0 / 16.0 GB + 2.6 GB RAM	Partial offload	up to 128k	Try IQ4_XS: Runs great thinking mode
Qwen3.5 27B	Runs slowly	est. 10.5 tok/s8.4–12.6calibrated estimate ±20%	Q4_K_M	16.0 / 16.0 GB + 2.8 GB RAM	Partial offload	up to 64k	Try IQ4_XS: Runs well UD-Q2_K_XL (heavy quality loss): Runs great thinking mode
Qwen3.8 27B	Runs slowly	est. 9.7 tok/s7.8–11.7calibrated estimate ±20%	Q4_K_M	16.0 / 16.0 GB + 3.2 GB RAM	Partial offload	up to 64k	Try IQ4_XS: Runs well Q2_K (heavy quality loss): Runs great thinking mode
Gemma 4 31B	Runs slowly	est. 7.9 tok/s6.3–9.5calibrated estimate ±20%	Q4_K_M	16.0 / 16.0 GB + 4.2 GB RAM	Partial offload	up to 64k	UD-Q2_K_XL (heavy quality loss): Runs great thinking mode
EXAONE 4.0 32B	Runs slowly	est. 6.9 tok/s5.5–8.2calibrated estimate ±20%	Q4_K_M	16.0 / 16.0 GB + 5.1 GB RAM	Partial offload	up to 64k	thinking mode
Granite 4.2 30B	Runs slowly	est. 6.4 tok/s5.2–7.7calibrated estimate ±20%	Q4_K_M	16.0 / 16.0 GB + 5.0 GB RAM	Partial offload	up to 16k	Q2_K (heavy quality loss): Runs great thinking mode
EXAONE 4.5 33B	Runs slowly	est. 6.2 tok/s4.9–7.4calibrated estimate ±20%	Q4_K_M	16.0 / 16.0 GB + 5.8 GB RAM	Partial offload	up to 64k	thinking mode
Qwen3 32B	Runs slowly	est. 4.9 tok/s3.9–5.9calibrated estimate ±20%	Q4_K_M	16.0 / 16.0 GB + 7.1 GB RAM	Partial offload	up to 16k	Q2_K (heavy quality loss): Runs great thinking mode
DeepSeek R1 Distill Qwen 32B	Runs slowly	est. 4.9 tok/s3.9–5.8calibrated estimate ±20%	Q4_K_M	16.0 / 16.0 GB + 7.2 GB RAM	Partial offload	up to 16k	Q2_K (heavy quality loss): Runs great thinking mode
Nanbeige4.3 33B	Runs slowly	est. 4.6 tok/s3.7–5.6calibrated estimate ±20%	Q4_K_M	16.0 / 16.0 GB + 7.7 GB RAM	Partial offload	up to 16k	Q2_K (heavy quality loss): Runs great thinking mode
Kolibri 1	Won't runTry Q2_K (heavy quality loss): Runs well	est. 65.5 tok/s45.8–85.1theoretical estimate ±30%	Q2_K	15.0 / 16.0 GB + 16.3 GB RAM	MoE experts in RAM	up to 256k	Not fully on the GPU — Runs great needs the whole model in GPU memory With 64 GB RAM: Runs well thinking mode
Solar Open 100B	Won't runTry Q2_K (heavy quality loss): Runs slowly	est. 19.6 tok/s13.7–25.5theoretical estimate ±30%	Q2_K	14.7 / 16.0 GB + 25.8 GB RAM	MoE experts in RAM	up to 16k	Experts run from system RAM — Runs well needs 30 tok/s or more on this path With 64 GB RAM: Runs slowly thinking mode
Llama 3.3 70B	Won't runTry Q2_K (heavy quality loss): Runs slowly	est. 2.6 tok/s2.1–3.2calibrated estimate ±20%	Q2_K	16.0 / 16.0 GB + 14.2 GB RAM	Partial offload	up to 16k	—
DeepSeek V4 Flash 0731	Won't run	—	MXFP4	needs 155.1 GB	—	—	Needs about 155.1 GB; 15.4 GB of VRAM and 28.0 GB of RAM are free thinking mode
GLM-5.3 Flash	Won't run	—	Q4_K_M	needs 181.4 GB	—	—	Needs about 181.4 GB; 15.4 GB of VRAM and 28.0 GB of RAM are free thinking mode
gpt-oss-120b	Won't run	—	MXFP4	needs 63.6 GB	—	—	Needs about 63.6 GB; 15.4 GB of VRAM and 28.0 GB of RAM are free With 64 GB RAM: Runs slowly thinking mode
Laguna S 2.1	Won't run	—	Q4_K_M	needs 68.5 GB	—	—	Needs about 68.5 GB; 15.4 GB of VRAM and 28.0 GB of RAM are free With 64 GB RAM: Runs slowly thinking mode
Ling-3.0 Flash	Won't run	—	Q4_K_M	needs 75.4 GB	—	—	Needs about 75.4 GB; 15.4 GB of VRAM and 28.0 GB of RAM are free With 96 GB RAM: Runs well thinking mode
Qwen3.5 122B-A10B	Won't run	—	Q4_K_M	needs 76.6 GB	—	—	Needs about 76.6 GB; 15.4 GB of VRAM and 28.0 GB of RAM are free With 96 GB RAM: Runs slowly thinking mode
Solar Open 2 250B	Won't run	—	IQ4_XS	needs 136.5 GB	—	—	Needs about 136.5 GB; 15.4 GB of VRAM and 28.0 GB of RAM are free thinking mode

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.

Measured results on the Radeon RX 9070

Public benchmarks we calibrate against. Their conditions (context, backend, flags) can differ from the estimates above.

Measured results on the Radeon RX 9070
Model	Quant	Backend	Context	Prompt (tok/s)	Generation (tok/s)	Flags	Source	Measured
llama-2-7b	Q4_0	llama.cpp	512	3,164.0	119.7	Vulkan	github.com	2025-08-17
Qwen3 8B	Q4_K_M	llama.cpp	512	2,635.0	109.9	Vulkan	github.com	2026-09-06
Qwen3 8B	Q4_K_M	llama.cpp	512	3,447.0	88.7	ROCm	github.com	2026-09-07

Frequently asked questions

What is the largest model that runs entirely on the Radeon RX 9070?

Muse Glimmer 30B at IQ4_XS (a 15.44 GB file) fits entirely in 16 GB with 8k context, at est. 31.5 tok/s (27.7–35.3, calibrated estimate ±12%).

How many local LLMs run well on the Radeon RX 9070?

At Q4_K_M with 8k context, out of 59 tracked models: 29 run great, 8 run well, 12 run slowly and 10 won't run.

Can the Radeon RX 9070 run a 70B model like Llama 3.3 70B?

Runs slowly — Partial offload, Q2_K (heavy quality loss): est. 2.6 tok/s (2.1–3.2, 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.