Skip to content

Qwen3.6 35B A3B VRAM requirements

How much VRAM Qwen3.6 35B A3B needs
4-bit weights: 20.3 GBWith 8K context: 21.5 GBWith 32K context: 22.0 GB8-bit, 8K context: 36.8 GB
Parameters
36B
Active per token
3B
Longest context
256K
Released
Apr 2026
License
Apache 2.0
Strongest at
Chat, Coding, Agents, Docs

How much VRAM Qwen3.6 35B A3B needs

Memory Qwen3.6 35B A3B needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit67.0 GB68.2 GB68.6 GB70.5 GB
8-bit35.6 GB36.8 GB37.2 GB39.1 GB
6-bit27.5 GB28.7 GB29.1 GB31.0 GB
5-bit23.8 GB25.0 GB25.5 GB27.4 GB
4-bit20.3 GB21.5 GB22.0 GB23.8 GB
3-bit16.4 GB17.6 GB18.0 GB19.9 GB
2-bit12.6 GB13.8 GB14.2 GB16.1 GB
1.6-bit6.9 GB8.1 GB8.6 GB10.5 GB

One conversation, 16-bit context cache, runtime overhead included.

GPUs and Macs that run Qwen3.6 35B A3B

The best way Qwen3.6 35B A3B runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: Qwen3.6 35B A3B on RTX 4060 report
RTX 5060Too bigPDF: Qwen3.6 35B A3B on RTX 5060 report
RTX 3070Too bigPDF: Qwen3.6 35B A3B on RTX 3070 report
RTX 3080 10 GBToo bigPDF: Qwen3.6 35B A3B on RTX 3080 10 GB report
RTX 3060 12 GBToo bigPDF: Qwen3.6 35B A3B on RTX 3060 12 GB report
Arc B580Too bigPDF: Qwen3.6 35B A3B on Arc B580 report
RTX 4070Too bigPDF: Qwen3.6 35B A3B on RTX 4070 report
RTX 4070 SuperToo bigPDF: Qwen3.6 35B A3B on RTX 4070 Super report
RTX 5070Too bigPDF: Qwen3.6 35B A3B on RTX 5070 report
RTX 4060 Ti 16 GBToo bigPDF: Qwen3.6 35B A3B on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GBToo bigPDF: Qwen3.6 35B A3B on RTX 5060 Ti 16 GB report
Arc A770 16 GBToo bigPDF: Qwen3.6 35B A3B on Arc A770 16 GB report
Radeon RX 9070 XTToo bigPDF: Qwen3.6 35B A3B on Radeon RX 9070 XT report
RTX 4070 Ti SuperToo bigPDF: Qwen3.6 35B A3B on RTX 4070 Ti Super report
RTX 4080 SuperToo bigPDF: Qwen3.6 35B A3B on RTX 4080 Super report
RTX 5070 TiToo bigPDF: Qwen3.6 35B A3B on RTX 5070 Ti report
RTX 5080Too bigPDF: Qwen3.6 35B A3B on RTX 5080 report
Radeon RX 7900 XTX4-bit72K146 tok/s0.7 sPDF: Qwen3.6 35B A3B on Radeon RX 7900 XTX report
RTX 30904-bit72K190 tok/s0.4 sPDF: Qwen3.6 35B A3B on RTX 3090 report
RTX 40904-bit72K197 tok/s0.2 sPDF: Qwen3.6 35B A3B on RTX 4090 report
Radeon AI PRO R97006-bit96K98 tok/s0.4 sPDF: Qwen3.6 35B A3B on Radeon AI PRO R9700 report
RTX 50906-bit96K227 tok/s0.1 sPDF: Qwen3.6 35B A3B on RTX 5090 report
2× RTX 30908-bit256K136 tok/s0.2 sPDF: Qwen3.6 35B A3B on 2× RTX 3090 report
2× RTX 40908-bit256K143 tok/s0.1 sPDF: Qwen3.6 35B A3B on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit256K171 tok/s0.2 sPDF: Qwen3.6 35B A3B on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit256K138 tok/s0.2 sPDF: Qwen3.6 35B A3B on RTX 6000 Ada report
L40S8-bit256K129 tok/s0.3 sPDF: Qwen3.6 35B A3B on L40S report
2× RTX 50908-bit256K200 tok/s0.1 sPDF: Qwen3.6 35B A3B on 2× RTX 5090 report
4× RTX 30908-bit256K136 tok/s0.1 sPDF: Qwen3.6 35B A3B on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit256K200 tok/s0.1 sPDF: Qwen3.6 35B A3B on RTX PRO 6000 Blackwell report
Mac mini M4 16 GBToo bigPDF: Qwen3.6 35B A3B on Mac mini M4 16 GB report
Mac mini M4 24 GBToo bigPDF: Qwen3.6 35B A3B on Mac mini M4 24 GB report
Mac mini M4 32 GB3-bit152K47 tok/s9.8 sPDF: Qwen3.6 35B A3B on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GBToo bigPDF: Qwen3.6 35B A3B on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB6-bit256K59 tok/s4.9 sPDF: Qwen3.6 35B A3B on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit256K50 tok/s4.9 sPDF: Qwen3.6 35B A3B on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit256K74 tok/s2.5 sPDF: Qwen3.6 35B A3B on M4 Max 128 GB report
M5 Max 128 GB8-bit256K79 tok/s0.7 sPDF: Qwen3.6 35B A3B on M5 Max 128 GB report
M3 Ultra 96 GB8-bit256K85 tok/s1.4 sPDF: Qwen3.6 35B A3B on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit256K85 tok/s1.4 sPDF: Qwen3.6 35B A3B on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit256K85 tok/s1.4 sPDF: Qwen3.6 35B A3B on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit256K102 tok/s0.4 sPDF: Qwen3.6 35B A3B on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit256K40 tok/s1.4 sPDF: Qwen3.6 35B A3B on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit256K40 tok/s1.4 sPDF: Qwen3.6 35B A3B on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit256K56 tok/s0.7 sPDF: Qwen3.6 35B A3B on RTX Spark report
DGX Spark8-bit256K52 tok/s0.8 sPDF: Qwen3.6 35B A3B on DGX Spark report
Dual DGX Spark8-bit256K52 tok/s0.5 sPDF: Qwen3.6 35B A3B on Dual DGX Spark report
DGX Station GB3008-bit256K327 tok/s0.1 sPDF: Qwen3.6 35B A3B on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit256K200 tok/s0.1 sPDF: Qwen3.6 35B A3B on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit256K200 tok/s0.1 sPDF: Qwen3.6 35B A3B on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit256K200 tok/s0.1 sPDF: Qwen3.6 35B A3B on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit256K214 tok/s0.3 sPDF: Qwen3.6 35B A3B on A100 80 GB SXM report
H100 SXM8-bit256K264 tok/s0.1 sPDF: Qwen3.6 35B A3B on H100 SXM report
4× H200 NVL8-bit256K297 tok/s0.1 sPDF: Qwen3.6 35B A3B on 4× H200 NVL report
8× H200 NVL8-bit256K297 tok/s0.1 sPDF: Qwen3.6 35B A3B on 8× H200 NVL report

Best fit is the highest quality, up to 8-bit, that leaves 16K of context. First reply is the wait for a 2K prompt. Speeds are rough, plus or minus 25%.

How to run it

Ollama
ollama run qwen3.6:35b-a3b
llama.cpp
llama-server -hf unsloth/Qwen3.6-35B-A3B-GGUF
vLLM
vllm serve Qwen/Qwen3.6-35B-A3B
MLX (Mac)
mlx_lm.chat --model mlx-community/Qwen3.6-35B-A3B-4bit

Questions

How much VRAM does Qwen3.6 35B A3B need?

Qwen3.6 35B A3B needs 20.3 GB for its weights at 4-bit, 21.5 GB with an 8K context and 22.0 GB with 32K. At 8-bit, 36.8 GB.

What is the minimum VRAM for Qwen3.6 35B A3B?

About 17.6 GB at 3-bit with an 8K context, the lowest precision we recommend for a model this size. The smallest machine here that runs it without spilling is the Radeon RX 7900 XTX.

How big is Qwen3.6 35B A3B?

36B parameters (3B active per token): 67.0 GB of weights at 16-bit, 20.3 GB at 4-bit.

Can Qwen3.6 35B A3B run on an RTX 4090?

Yes. At 4-bit it holds 72K of context and writes about 197 tokens a second.

More models