Will it run?

Which models fit which card.

50 open-weight models against the 17 cards in our catalog. Weights are sized from the parameter count (all experts count for MoE), a 1.5 GiB reserve covers the runtime and a small KV cache, and what is left is your context headroom. Tap any cell for the full verdict at Q4, Q8 and FP16.

The grid

Q4_K_M verdicts

fits comfortablytightCPU offloadno

Model · weights at Q4	RTX 509032GB	RTX 3090 (used)24GB	RTX 4090 (used)24GB	RX 7900 XTX24GB	RTX 5060 Ti 16GB16GB	RTX 5070 Ti16GB	RTX 508016GB	RX 7800 XT16GB	RX 9060 XT 16GB16GB	RX 907016GB	RX 9070 XT16GB	Arc B58012GB	RTX 3060 12GB (used)12GB	RTX 4070 Super12GB	RTX 507012GB	Arc B57010GB	RTX 50608GB
Qwen3 0.6B0.6B · 0.3 GiB	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Gemma 3 1B IT1B · 0.6 GiB	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Llama 3.2 1B Instruct1.23B · 0.7 GiB	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Qwen3 1.7B1.7B · 0.9 GiB	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Llama 3.2 3B Instruct3.21B · 1.8 GiB	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Phi-3.5 mini Instruct3.82B · 2.1 GiB	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Phi-4 mini Instruct3.84B · 2.1 GiB	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Qwen3 4B4B · 2.2 GiB	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Gemma 3 4B IT4.3B · 2.4 GiB	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Qwen3.5 4B4.66B · 2.6 GiB	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Mistral 7B Instruct v0.37.25B · 4.1 GiB	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Qwen2.5 7B Instruct7.61B · 4.3 GiB	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Qwen2.5 Coder 7B Instruct7.61B · 4.3 GiB	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
DeepSeek R1 Distill Qwen 7B7.62B · 4.3 GiB	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
DeepSeek R1 Distill Llama 8B8.03B · 4.5 GiB	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Hermes 3 Llama 3.1 8B8.03B · 4.5 GiB	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Llama 3.1 8B Instruct8.03B · 4.5 GiB	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Qwen3 8B8.2B · 4.6 GiB	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Qwen2.5-VL 7B Instruct8.29B · 4.6 GiB	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Gemma 2 9B IT9.24B · 5.2 GiB	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Qwen3.5 9B9.65B · 5.4 GiB	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Tight
Gemma 3 12B IT12.2B · 6.8 GiB	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Offload
Mistral NeMo 12B Instruct (2407)12.2B · 6.8 GiB	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Offload
Phi-4 (14B)14.7B · 8.2 GiB	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Tight	Offload
DeepSeek R1 Distill Qwen 14B14.8B · 8.3 GiB	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Tight	Offload
Qwen3 14B14.8B · 8.3 GiB	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Tight	Offload
gpt-oss-20b21B MoE (3.6B active) · 11.7 GiB	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Offload	Offload	Offload	Offload	Offload	Offload
Codestral 22B v0.122.2B · 12.4 GiB	Yes	Yes	Yes	Yes	Tight	Tight	Tight	Tight	Tight	Tight	Tight	Offload	Offload	Offload	Offload	Offload	Offload
Mistral Small 3.2 24B Instruct (2506)24B · 13.4 GiB	Yes	Yes	Yes	Yes	Tight	Tight	Tight	Tight	Tight	Tight	Tight	Offload	Offload	Offload	Offload	Offload	Offload
Gemma 4 26B-A4B IT25.8B MoE (3.8B active) · 14.4 GiB	Yes	Yes	Yes	Yes	Tight	Tight	Tight	Tight	Tight	Tight	Tight	Offload	Offload	Offload	Offload	Offload	Offload
Gemma 3 27B IT27.4B · 15.3 GiB	Yes	Yes	Yes	Yes	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	No
Qwen3.6 27B27.8B · 15.5 GiB	Yes	Yes	Yes	Yes	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	No
GLM-4.7-Flash30B MoE (3B active) · 16.8 GiB	Yes	Yes	Yes	Yes	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	No
Qwen3 30B-A3B30.5B MoE (3.3B active) · 17.0 GiB	Yes	Yes	Yes	Yes	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	No
Gemma 4 31B IT31.3B · 17.5 GiB	Yes	Yes	Yes	Yes	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	No
Qwen2.5 Coder 32B Instruct32.5B · 18.2 GiB	Yes	Yes	Yes	Yes	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	No
DeepSeek R1 Distill Qwen 32B32.8B · 18.3 GiB	Yes	Yes	Yes	Yes	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	No
Qwen3 32B32.8B · 18.3 GiB	Yes	Yes	Yes	Yes	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	No
Qwen3.6 35B-A3B36B MoE (3B active) · 20.1 GiB	Yes	Tight	Tight	Tight	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	No	No
Mixtral 8x7B Instruct v0.146.7B MoE (12.9B active) · 26.1 GiB	Tight	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	Offload	No	No	No	No	No	No
Llama 3.3 70B Instruct70.6B · 39.5 GiB	Offload	Offload	Offload	Offload	No	No	No	No	No	No	No	No	No	No	No	No	No
GLM-4.5-Air106B MoE (12B active) · 59.2 GiB	Offload	No	No	No	No	No	No	No	No	No	No	No	No	No	No	No	No
Llama 4 Scout (17B x 16E) Instruct109B MoE (17B active) · 60.9 GiB	Offload	No	No	No	No	No	No	No	No	No	No	No	No	No	No	No	No
gpt-oss-120b117B MoE (5.1B active) · 65.4 GiB	No	No	No	No	No	No	No	No	No	No	No	No	No	No	No	No	No
Mixtral 8x22B Instruct v0.1141B MoE (39B active) · 78.8 GiB	No	No	No	No	No	No	No	No	No	No	No	No	No	No	No	No	No
Qwen3 235B-A22B235B MoE (22B active) · 131.3 GiB	No	No	No	No	No	No	No	No	No	No	No	No	No	No	No	No	No
DeepSeek V4 Flash284B MoE (13B active) · 158.7 GiB	No	No	No	No	No	No	No	No	No	No	No	No	No	No	No	No	No
Llama 4 Maverick (17B x 128E) Instruct400B MoE (17B active) · 223.5 GiB	No	No	No	No	No	No	No	No	No	No	No	No	No	No	No	No	No
DeepSeek R1 (671B)671B MoE (37B active) · 375.0 GiB	No	No	No	No	No	No	No	No	No	No	No	No	No	No	No	No	No
Kimi K2 Instruct1000B MoE (32B active) · 558.8 GiB	No	No	No	No	No	No	No	No	No	No	No	No	No	No	No	No	No

Method: weights in GiB = parameters x bytes per parameter (Q4_K_M 0.60, Q8_0 1.06, FP16 2.0), plus a fixed 1.5 GiB reserve. "Fits comfortably" means that total sits within 85% of the card's memory; "tight" means it fits with little room for context; "CPU offload" means the weights exceed VRAM by up to 2x and can still run slowly with layers offloaded to system RAM. Parameter counts come from each model's card or release notes; the per-token KV cache depends on the architecture and is not in this table, so treat the headroom figures as a budget, not a promise.

Which models fit which card.

The grid

The Silicon Note