AMD / discrete

AMD Radeon RX 9070

16 GB decides what fits. 640 GB/s decides how fast it runs once it does.

Open in the calculator →Best models for 16 GB, on every card that size →

Find one: Amazon ↗ · eBay (new and used) ↗Store links may pay us a commission. They never decide which card is suggested — the memory arithmetic does.

Quick answer

The AMD Radeon RX 9070 fits 153 of 320 sized models entirely in its 16 GB; the largest widely used one is gpt-oss-20b (13.8 GB at Q4_K_M). Memory decides what fits; its 640 GB/s of bandwidth decides how fast it answers.

Largest popular model that fitsgpt-oss-20b21B params · Q4_K_M · needs 13.8 GB~199 tok/s Faster than you readOpen in the calculator →
Best fast pickERNIE-4.5-21B-A3B-PT22B params · Q4_K_M · needs 14.7 GB~196 tok/s Faster than you readOpen in the calculator →
Most downloaded that fitsQwen3.5-9B9.7B params · Q4_K_M · needs 7.05 GB~103 tok/s Faster than you readOpen in the calculator →

At 8,192 tokens of context with the whole model in device memory. Speeds are estimates from memory bandwidth, not benchmarks run on this card.

AMD Radeon RX 9070 · 16 GB · 640 GB/s · Q4_K_M where published · 8,192 tokensSpecificationpublished dataDecode speedestimate

153 of 320 fit entirely

153 of the 320 models the engine can size fit entirely in device memory at Q4_K_M where it is published, otherwise the nearest published format, and 8,192 tokens. 73 run with some layers on system memory (32 GB assumed), and 94 do not run at all.

48%of the models the engine can size fit entirely

16GB of device memory

640GB/s memory bandwidth

73run with system memory

94do not run at 8,192 tokens

Featured models · Q4_K_M at 8,192 tokens, including offload

Model	Needs	Verdict	Decode	Calculator
Qwen3.5-2B2.3B parameters	2.32 GB	fits	~404 tok/s	Open →
NVIDIA-Nemotron-3-Nano-4B-BF164.0B parameters	3.46 GB	fits	~187 tok/s	Open →
Qwen3-VL-8B-Instruct8.8B parameters	7.04 GB	fits	~85 tok/s	Open →
Qwen3.5-9B9.7B parameters	7.05 GB	fits	~103 tok/s	Open →
gemma-4-12B-it12B parameters	9.54 GB	fits	~61 tok/s	Open →
gpt-oss-20b21B parameters	13.8 GB	fits	~199 tok/s	Open →
Qwen3.8-27B28B parameters	18.6 GB	10 of 64 layers on system RAM (2.66 GB)	~16 tok/s with offload	Open →
Qwen3.6-27B28B parameters	18.6 GB	10 of 64 layers on system RAM (2.66 GB)	~16 tok/s with offload	Open →
Kimi-Linear-48B-A3B-Instruct49B parameters	30.4 GB	14 of 27 layers on system RAM (15.3 GB)	~59 tok/s with offload	Open →
Qwen3.8-Flash-Next180B parameters	111.6 GB	needs 96.7 GB of system RAM; 32 GB assumed	does not run	Open →
DeepSeek-V4-Flash-Vision-Exp305B parameters	188.1 GB	needs 174.2 GB of system RAM; 32 GB assumed	does not run	Open →
GLM-5.3-Flash321B parameters	198.3 GB	needs 184.1 GB of system RAM; 32 GB assumed	does not run	Open →

These examples are selected from prominent labs using the catalogue’s latest Hugging Face 30-day downloads and repository-creation freshness signal, with newer releases guaranteed a place. Offloaded rows assume 32 GB of system RAM, and a speed is only shown for a row that runs. Fitting in memory is not the same as loading: whether the runtime and version you have supports each architecture and format on this machine has not been tested here. Each “Open” link carries the same model, format, context and RAM into the calculator.

Fits entirely in AMD Radeon RX 9070 memory — 153 of 320 sized models

Every model that fitswhole model resident · offload off

The most demanding model that fits is gpt-neox-20b at Q4_K_M: 15.7 GB of the 16 GB, leaving 0.25 GB spare.

Fully resident at 8,192 tokens (or the model’s own maximum, where that is shorter), offload off, at 640 GB/s. Each row is a run of the engine for this configuration; the rows start with current, prominent releases and “fits” is memory, not a tested runtime. Older or less prominent models remain available through this search and “Show all”.

Find a model in this list

Showing 40 of 153 models that fit.

Model	Format	Needs	Spare	Decode	Downloads / 30d	Calculator
Qwen3.5-9BQwen · 9.7B params	Q4_K_M	7.05 GB	8.95 GB	~103 tok/s	9M	Open →
Qwen3.5-4BQwen · 4.7B params	Q4_K_M	3.98 GB	12.0 GB	~185 tok/s	7.8M	Open →
gemma-4-12B-itGoogle · 12B params	Q4_K_M	9.54 GB	6.46 GB	~61 tok/s	1.9M	Open →
Qwen3.5-2BQwen · 2.3B params	Q4_K_M	2.32 GB	13.7 GB	~404 tok/s	4.9M	Open →
gemma-4-E4B-itGoogle · 8.0B params	Q4_K_M	5.86 GB	10.1 GB	~165 tok/s	4.4M	Open →
NVIDIA-Nemotron-3-Nano-4B-BF16NVIDIA · 4.0B params	Q4_K_M	3.46 GB	12.5 GB	~187 tok/s	3.4M	Open →
gemma-4-E2B-itGoogle · 5.1B params	Q4_K_M	4.02 GB	12.0 GB	~435 tok/s	3M	Open →
Qwen3.5-0.8BQwen · 873M params	Q4_K_M	1.46 GB	14.5 GB	~886 tok/s	2.6M	Open →
Qwen3-VL-8B-InstructQwen · 8.8B params	Q4_K_M	7.04 GB	8.96 GB	~85 tok/s	14.6M	Open →
North-Micro-Vision-InstructCohereLabs · 2.5B params	Q4_K_M	2.99 GB	13.0 GB	~266 tok/s	180.7K	Open →
granite-4.2-8bIBM · 8.8B params	Q4_K_M	7.49 GB	8.51 GB	~77 tok/s	128.4K	Open →
granite-4.1-3bIBM · 3.4B params	Q4_K_M	3.72 GB	12.3 GB	~180 tok/s	521.8K	Open →
LFM2.5-2.6BLiquidAI · 2.7B params	Q4_K_M	2.61 GB	13.4 GB	~289 tok/s	108.1K	Open →
Qwen3-VL-4B-InstructQwen · 4.4B params	Q4_K_M	4.72 GB	11.3 GB	~135 tok/s	3.5M	Open →
Qwen3-VL-2B-InstructQwen · 2.1B params	Q4_K_M	3.05 GB	13.0 GB	~244 tok/s	2.8M	Open →
granite-4.2-3bIBM · 3.7B params	Q4_K_M	3.72 GB	12.3 GB	~180 tok/s	50.6K	Open →
LFM2.5-230MLiquidAI · 230M params	Q4_K_M	1.05 GB	14.9 GB	~2013 tok/s	88.6K	Open →
Qwen3-0.6BQwen · 752M params	Q4_K_M	2.22 GB	13.8 GB	~374 tok/s	29.7M	Open →
gpt-oss-20bOpenAI · 21B params	Q4_K_M	13.8 GB	2.24 GB	~199 tok/s	6.6M	Open →
granite-4.1-8bIBM · 8.8B params	Q4_K_M	7.49 GB	8.51 GB	~77 tok/s	179.3K	Open →
LFM2.5-VL-3BLiquidAI · 3.1B params	Q4_K_M	2.85 GB	13.1 GB	~289 tok/s	26.7K	Open →
MiMo-V2.6-Distill-Qwen-9BNEWXiaomiMiMo · 9.4B params	Q4_K_M	6.90 GB	9.10 GB	~103 tok/s	14.6K	Open →
Qwen3-4B-Instruct-2507Qwen · 4.0B params	Q4_K_M	4.72 GB	11.3 GB	~135 tok/s	3.7M	Open →
Ling-3.0-tinyinclusionAI · 7.9B params	Q4_K_M	5.85 GB	10.1 GB	~581 tok/s	17.7K	Open →
Qwen3-8BQwen · 8.2B params	Q4_K_M	7.04 GB	8.96 GB	~85 tok/s	10.7M	Open →
Olmo-3-7B-Instructallenai · 7.3B params	Q4_K_M	7.96 GB	8.04 GB	~72 tok/s	481.7K	Open →
Qwen3-4BQwen · 4.0B params	Q4_K_M	4.51 GB	11.5 GB	~135 tok/s	7.8M	Open →
LFM2.5-8B-A1BLiquidAI · 8.5B params	Q4_K_M	6.06 GB	9.94 GB	~450 tok/s	32.4K	Open →
LFM2.5-350MLiquidAI · 354M params	Q4_K_M	1.13 GB	14.9 GB	~1299 tok/s	69.9K	Open →
Hy-MT2-1.8BTencent · 2.0B params	Q4_K_M	2.47 GB	13.5 GB	~299 tok/s	28.9K	Open →
LLaDA2.0-miniinclusionAI · 16B params	Q4_K_M	11.0 GB	5.00 GB	~466 tok/s	217.2K	Open →
LFM2.5-1.2B-InstructLiquidAI · 1.2B params	Q4_K_M	1.63 GB	14.4 GB	~479 tok/s	119.2K	Open →
Ministral-3-14B-Instruct-2512Mistral AI · 14B params	Q4_K_M	10.4 GB	5.62 GB	~54 tok/s	251K	Open →
Qwen3-1.7BQwen · 2.0B params	Q4_K_M	3.02 GB	13.0 GB	~244 tok/s	3.1M	Open →
Nemotron-3.5-Content-SafetyNVIDIA · 4.3B params	Q4_K_M	3.73 GB	12.3 GB	~180 tok/s	14.1K	Open →
Hy-MT2-7BTencent · 8.0B params	Q4_K_M	6.50 GB	9.50 GB	~88 tok/s	15.4K	Open →
Qwen3-14BQwen · 15B params	Q4_K_M	11.1 GB	4.86 GB	~50 tok/s	2.7M	Open →
GLM-4.6V-Flashzai-org · 10B params	Q4_K_M	7.45 GB	8.55 GB	~87 tok/s	103.7K	Open →
Qwen2.5-VL-7B-InstructQwen · 8.3B params	Q4_K_M	6.36 GB	9.64 GB	~103 tok/s	5.8M	Open →
SmolLM3-3BHuggingFaceTB · 3.1B params	Q4_K_M	3.46 GB	12.5 GB	~198 tok/s	617K	Open →

Sized at the format most people actually download, not at FP16. The catalogue ordered by downloads →

Find the AMD Radeon RX 9070: Amazon ↗ · eBay (new and used) ↗Store links may pay us a commission. They never decide which card is suggested — the memory arithmetic does.

Under the hood

The specification behind every figure

What the manufacturer publishes for this device, and the pages it was read from.

Manufacturer specification

Memory scope	dedicated
Capacity	16 GB
Published options	16 GB
Bandwidth	640 GB/s
Memory type	GDDR6
Bus width	256-bit
FP32 peak	36.1 TFLOPS
Dense matrix peak	145 TFLOPS (FP16 Matrix, dense)
TBP	220 W

Source ledger

Radeon RX 9070 ↗

Retrieved 2026-09-12

Caveats

AMD publishes bandwidth as "Up to 640 GB/s"; that ceiling is the figure used, and it is the same arithmetic as the published 256-bit interface at 20 Gbps.

Published capacity is a hardware ceiling, not guaranteed free runtime memory. The calculator shows the runtime reserve separately rather than folding it into a single number.

Run the diagnostic on the AMD Radeon RX 9070 →