Can I run Qwen3-Coder-Next on an RTX 5090?

With --n-cpu-moe 17: the Q4_K_M GGUF keeps 31.3 GB on the RTX 5090 and 15.7 GB in system RAM, with the experts of 17 of its 48 layers moved there. Whole, Qwen3-Coder-Next needs about 50.7 GB at Q4_K_M with 32K tokens of context, 18.7 GB more than the 32 GB RTX 5090 holds. The smallest setup here that holds Qwen3-Coder-Next at Q4_K_M with 32K is 2× RTX 5090 (64 GB).

With offload the Q4_K_M GGUF with --n-cpu-moe 17 and 32K tokens of context

On the card: 31.3 GB
RTX 5090: 32 GB, 1,792 GB/s
In RAM: 15.7 GB
Tokens/s: 52–91 tokens/s

With the Q4_K_M GGUF, --n-cpu-moe 17 and 32K tokens of context it writes about 52–91 tokens/s for one request on an RTX 5090 and dual-channel DDR5-5600.

Qwen3-Coder-Next on the RTX 5090 as the context fills

One request, FP16 KV cache, 0.5 GB plus 10% overhead; a minus sign is memory missing, and tight is less than 0.5 GB free.

Context	Q4_K_M	Free	Tokens/s	Q8_0	Free	Tokens/s
4K	50.0 GB	−18.0 GB	—	87.3 GB	−55.3 GB	—
8K	50.1 GB	−18.1 GB	—	87.4 GB	−55.4 GB	—
16K	50.3 GB	−18.3 GB	—	87.6 GB	−55.6 GB	—
32K	50.7 GB	−18.7 GB	—	88.0 GB	−56.0 GB	—
64K	51.5 GB	−19.5 GB	—	88.9 GB	−56.9 GB	—
128K	53.2 GB	−21.2 GB	—	90.5 GB	−58.5 GB	—
256K (full)	56.5 GB	−24.5 GB	—	93.8 GB	−61.8 GB	—

--n-cpu-moe for Qwen3-Coder-Next on the RTX 5090

With --n-cpu-moe 17 the Q4_K_M GGUF keeps 31.3 GB on the RTX 5090 and 15.7 GB in RAM, about 52–91 tokens/s with DDR5-5600. Measured Q4_K_M file, 1 GB of buffers; tokens/s by system RAM speed.

Context	--n-cpu-moe	On the card	In RAM	DDR4-3200	DDR5-5600	DDR5-6400
8K	16	31.6 GB	14.8 GB	39–66	58–100	63–110
32K	17	31.3 GB	15.7 GB	36–61	52–91	57–99
64K	18	31.1 GB	16.6 GB	32–55	47–81	51–88
128K	19	31.7 GB	17.5 GB	29–49	40–69	43–74
256K (full)	23	31.2 GB	21.0 GB	22–37	30–51	32–54

Qwen3-Coder-Next on more than one RTX 5090

Cards	Best at 32K	Tokens/s	Q4_K_M longest	Q4_K_M tokens/s	Rent per hour
2× (64 GB)	Q5_K_M	161–341	256K (full)	168–362	$1.38
4× (128 GB)	Q8_0	187–414	256K (full)	212–492	$2.76

Tensor parallel with the combined bandwidth, as in vLLM; llama.cpp splits layers by default and runs at about one card's speed.

Other options

The next smaller setting, Q3_K_M, takes 41.2 GB at 32K, 9.22 GB over the RTX 5090; it does not fit with 0.5 GB to spare even with 1K tokens.
Smallest setup for Q4_K_M at 32K: 2× RTX 5090 (64 GB).

Run Qwen3-Coder-Next on the RTX 5090 with llama-server

llama-server -hf unsloth/Qwen3-Coder-Next-GGUF:Q4_K_M -c 32768 --n-cpu-moe 17

Qwen3-Coder-Next-Q4_K_M.gguf, 48.5 GB, from unsloth/Qwen3-Coder-Next-GGUF (checked 2026-09-29); --n-cpu-moe 17 with -c 32768 is the plan above.

Questions

Can I run Qwen3-Coder-Next on an RTX 5090?

How fast is Qwen3-Coder-Next on an RTX 5090?

With the Q4_K_M GGUF, --n-cpu-moe 17 and 32K tokens of context it writes about 52–91 tokens/s for one request on an RTX 5090 and dual-channel DDR5-5600.

Does Qwen3-Coder-Next need --n-cpu-moe on an RTX 5090?

With --n-cpu-moe 17 the Q4_K_M GGUF keeps 31.3 GB on the RTX 5090 and 15.7 GB in RAM, about 52–91 tokens/s with DDR5-5600.

What does a second RTX 5090 change for Qwen3-Coder-Next?

Two RTX 5090 cards (64 GB in one tensor-parallel group) hold Qwen3-Coder-Next at Q5_K_M with 32K, and Q4_K_M up to 256K (full) tokens, at about 168–362 tokens/s.

Try other settings in the VRAM calculator, the speed calculator or the MoE offload planner. See also Qwen3-Coder-Next VRAM requirements, what LLMs an RTX 5090 can run and every pair, or detect your own GPU. Model data checked 2026-09-26.