You downloaded the model.
Now what?

Tell us what is in your computer. We will tell you what you can run — and whether it will actually be useful, which is not the same question.

VRAM calculators answer one question: does it fit? And they answer it wrongly for MoE models, because they do not know the experts can spill into system memory. That is why they told you „not possible" and you gave up.

.fisa, #rezultate, #formula { display: none; } The advisor needs JavaScript, because the arithmetic runs in your browser — our server generates manuals and is not allowed to be busy with this.

But do not leave empty-handed. The most asked answer: yes, a 35-billion-parameter model runs on a card from 2017 with 11 GB, at 23.7 tokens per second — measured by us over seven hours on more than 500 real answers. It works because the experts of an MoE can live in system memory, and VRAM calculators do not know that, which is why they tell you „not possible".

The guide explains it all · FAQ · For machines · How we measure

What is in your computer

Graphics card

System RAM (GB)

Memory type

Quantisation

What you can run

Below is the same answer, worked out in advance for six common machines, with 32 GB of RAM and the usual Q4 compression. It is here in plain HTML on purpose: a program reading this page should not have to run our JavaScript to learn anything. Use the controls above for your own machine.

No graphics card (CPU only) 13 of 14 models run

Model	Runs?	Speed	Label
Llama 3.3 70B	no	—	NOT POSSIBLE
Qwen3.6 35B-A3B	yes	~7.1 tok/s	ESTIMATED
Qwen3 Coder 30B-A3B	yes	~7.1 tok/s	ESTIMATED
Qwen 27B (dense)	yes	~0.8 tok/s	ESTIMATED
Gemma 3 27B	yes	~0.8 tok/s	ESTIMATED
Mistral Small 24B	yes	~0.9 tok/s	ESTIMATED
GPT-OSS 20B	yes	~5.9 tok/s	ESTIMATED
Phi-4 14B	yes	~1.5 tok/s	ESTIMATED
Qwen 14B	yes	~1.5 tok/s	ESTIMATED
Gemma 3 12B	yes	~1.8 tok/s	ESTIMATED
Llama 3 8B	yes	~2.7 tok/s	ESTIMATED
Mistral 7B	yes	~3.0 tok/s	ESTIMATED
Qwen 4B (small brain)	yes	~5.3 tok/s	ESTIMATED
Qwen 1.7B	yes	~12.5 tok/s	ESTIMATED

GTX 1050 Ti 12 of 14 models run

Model	Runs?	Speed	Label
Llama 3.3 70B	no	—	NOT POSSIBLE
Qwen3.6 35B-A3B	no	—	NOT POSSIBLE
Qwen3 Coder 30B-A3B	yes	~17.0 tok/s	ESTIMATED
Qwen 27B (dense)	yes, slowly	0.6–0.8 tok/s	ESTIMATED
Gemma 3 27B	yes, slowly	0.6–0.8 tok/s	ESTIMATED
Mistral Small 24B	yes, slowly	0.7–0.9 tok/s	ESTIMATED
GPT-OSS 20B	yes	~14.2 tok/s	ESTIMATED
Phi-4 14B	yes, slowly	1.2–1.7 tok/s	ESTIMATED
Qwen 14B	yes, slowly	1.2–1.7 tok/s	ESTIMATED
Gemma 3 12B	yes, slowly	1.4–2.0 tok/s	ESTIMATED
Llama 3 8B	yes, slowly	2.3–3.2 tok/s	ESTIMATED
Mistral 7B	yes, slowly	2.7–3.8 tok/s	ESTIMATED
Qwen 4B (small brain)	yes, slowly	5.8–8.3 tok/s	ESTIMATED
Qwen 1.7B	yes	49.4–71.4 tok/s	ESTIMATED

GTX 1080 Ti 13 of 14 models run

Model	Runs?	Speed	Label
Llama 3.3 70B	no	—	NOT POSSIBLE
Qwen3.6 35B-A3B	yes	~17.0 tok/s	ESTIMATED
Qwen3 Coder 30B-A3B	yes	~17.0 tok/s	ESTIMATED
Qwen 27B (dense)	yes, slowly	1.0–1.5 tok/s	ESTIMATED
Gemma 3 27B	yes, slowly	1.0–1.5 tok/s	ESTIMATED
Mistral Small 24B	yes, slowly	1.3–1.9 tok/s	ESTIMATED
GPT-OSS 20B	yes	~14.2 tok/s	ESTIMATED
Phi-4 14B	yes	25.9–37.5 tok/s	ESTIMATED
Qwen 14B	yes	25.9–37.5 tok/s	ESTIMATED
Gemma 3 12B	yes	30.3–43.7 tok/s	ESTIMATED
Llama 3 8B	yes	45.4–65.5 tok/s	ESTIMATED
Mistral 7B	yes	51.9–74.9 tok/s	ESTIMATED
Qwen 4B (small brain)	yes	90.8–120.0 tok/s	ESTIMATED
Qwen 1.7B	yes	over 120 tok/s	ESTIMATED

RTX 3060 12GB 14 of 14 models run

Model	Runs?	Speed	Label
Llama 3.3 70B	yes, slowly	0.3–0.4 tok/s	ESTIMATED
Qwen3.6 35B-A3B	yes	~17.0 tok/s	ESTIMATED
Qwen3 Coder 30B-A3B	yes	~17.0 tok/s	ESTIMATED
Qwen 27B (dense)	yes, slowly	1.1–1.6 tok/s	ESTIMATED
Gemma 3 27B	yes, slowly	1.1–1.6 tok/s	ESTIMATED
Mistral Small 24B	yes, slowly	1.4–2.1 tok/s	ESTIMATED
GPT-OSS 20B	yes	~14.2 tok/s	ESTIMATED
Phi-4 14B	yes	19.3–27.9 tok/s	ESTIMATED
Qwen 14B	yes	19.3–27.9 tok/s	ESTIMATED
Gemma 3 12B	yes	22.5–32.5 tok/s	ESTIMATED
Llama 3 8B	yes	33.8–48.8 tok/s	ESTIMATED
Mistral 7B	yes	38.6–55.7 tok/s	ESTIMATED
Qwen 4B (small brain)	yes	67.5–97.5 tok/s	ESTIMATED
Qwen 1.7B	yes	over 120 tok/s	ESTIMATED

RTX 4070 14 of 14 models run

Model	Runs?	Speed	Label
Llama 3.3 70B	yes, slowly	0.3–0.4 tok/s	ESTIMATED
Qwen3.6 35B-A3B	yes	~17.0 tok/s	ESTIMATED
Qwen3 Coder 30B-A3B	yes	~17.0 tok/s	ESTIMATED
Qwen 27B (dense)	yes, slowly	1.2–1.7 tok/s	ESTIMATED
Gemma 3 27B	yes, slowly	1.2–1.7 tok/s	ESTIMATED
Mistral Small 24B	yes, slowly	1.5–2.2 tok/s	ESTIMATED
GPT-OSS 20B	yes	~14.2 tok/s	ESTIMATED
Phi-4 14B	yes	27.0–39.0 tok/s	ESTIMATED
Qwen 14B	yes	27.0–39.0 tok/s	ESTIMATED
Gemma 3 12B	yes	31.5–45.5 tok/s	ESTIMATED
Llama 3 8B	yes	47.3–68.3 tok/s	ESTIMATED
Mistral 7B	yes	54.0–78.0 tok/s	ESTIMATED
Qwen 4B (small brain)	yes	94.5–120.0 tok/s	ESTIMATED
Qwen 1.7B	yes	over 120 tok/s	ESTIMATED

RTX 4090 14 of 14 models run

Model	Runs?	Speed	Label
Llama 3.3 70B	yes, slowly	0.4–0.6 tok/s	ESTIMATED
Qwen3.6 35B-A3B	yes	over 120 tok/s	ESTIMATED
Qwen3 Coder 30B-A3B	yes	over 120 tok/s	ESTIMATED
Qwen 27B (dense)	yes	28.0–40.4 tok/s	ESTIMATED
Gemma 3 27B	yes	28.0–40.4 tok/s	ESTIMATED
Mistral Small 24B	yes	31.5–45.5 tok/s	ESTIMATED
GPT-OSS 20B	yes	over 120 tok/s	ESTIMATED
Phi-4 14B	yes	54.0–78.0 tok/s	ESTIMATED
Qwen 14B	yes	54.0–78.0 tok/s	ESTIMATED
Gemma 3 12B	yes	63.0–91.0 tok/s	ESTIMATED
Llama 3 8B	yes	94.5–120.0 tok/s	ESTIMATED
Mistral 7B	yes	108.0–120.0 tok/s	ESTIMATED
Qwen 4B (small brain)	yes	over 120 tok/s	ESTIMATED
Qwen 1.7B	yes	over 120 tok/s	ESTIMATED

MEASURED means it happened, on a named machine, on a named date. ESTIMATED is computed from a measurement, and the formula is published. A model marked NOT POSSIBLE does not fit; one marked yes, slowly is split between the card and system memory, which works but reads everything again for every word.

The MEASURED and ESTIMATED labels are not decoration. Here is exactly how we measure — and why every input set contains one case that must fail.

Why trust us

Because the numbers labelled MEASURED are not derived from spec-sheet formulas — they happened, on a specific machine, on a specific date, and both are written next to them. Everything else honestly says ESTIMATED, with the formula in plain sight. And where we do not know, we put no number at all.

Since late August we have kept a 35-billion-parameter model alive, in production, on a graphics card from 2017 — which is how we know what happens after six hours, not just in the first minute.

And if the answer is yes?

guide