Alibaba · 35B · UD-Q4_K_M · 20.6 GiB

Qwen3.6 35B-A3B at UD-Q4_K_M needs about 22.4 GiB at 8K context, and the smallest listed machine that fits it is GeForce RTX 5090.

Hybrid MoE: 256 experts, 8 active, plus a shared expert. 30 layers are linear attention and 10 are full attention.

Weights: Qwen3.6-35B-A3B-UD-Q4_K_M.gguf. Architecture: config source. Hugging Face repo created 2026-04-15. License field: apache-2.0.

Quants with a measured file size

QuantFileSize
UD-Q4_K_M · recommendedQwen3.6-35B-A3B-UD-Q4_K_M.gguf20.6 GiB
UD-Q5_K_MQwen3.6-35B-A3B-UD-Q5_K_M.gguf24.6 GiB
UD-Q6_KQwen3.6-35B-A3B-UD-Q6_K.gguf27.3 GiB
Q8_0Qwen3.6-35B-A3B-Q8_0.gguf34.4 GiB

Machines

Same quant and 8,192 tokens on every row.
MachineMemoryTotalVerdictSpeed
Apple M1, 8 GB8 GB unified22.4 GiBNo—
Apple M2, 8 GB8 GB unified22.4 GiBNo—
Apple M3, 8 GB8 GB unified22.4 GiBNo—
GeForce RTX 3060 Ti8 GB GDDR22.4 GiBOffload9–13 tok/s
GeForce RTX 30708 GB GDDR22.4 GiBOffload9–13 tok/s
GeForce RTX 40608 GB GDDR22.4 GiBOffload9–13 tok/s
GeForce RTX 4060 Ti 8GB8 GB GDDR22.4 GiBOffload9–13 tok/s
GeForce RTX 50608 GB GDDR22.4 GiBOffload18–26 tok/s
GeForce RTX 5060 Ti 8GB8 GB GDDR22.4 GiBOffload18–26 tok/s
Radeon RX 66008 GB GDDR622.4 GiBOffload9–13 tok/s
Radeon RX 76008 GB GDDR622.4 GiBOffload9–13 tok/s
GeForce RTX 3080 10GB10 GB GDDR22.4 GiBOffload11–16 tok/s
GeForce RTX 3060 12GB12 GB GDDR22.4 GiBOffload12–18 tok/s
GeForce RTX 407012 GB GDDR22.4 GiBOffload13–18 tok/s
GeForce RTX 4070 Super12 GB GDDR22.4 GiBOffload13–18 tok/s
GeForce RTX 4070 Ti12 GB GDDR22.4 GiBOffload13–18 tok/s
GeForce RTX 507012 GB GDDR22.4 GiBOffload25–36 tok/s
Intel Arc B58012 GB GDDR622.4 GiBOffload13–18 tok/s
Radeon RX 6700 XT12 GB GDDR622.4 GiBOffload12–18 tok/s
Radeon RX 7700 XT12 GB GDDR622.4 GiBOffload13–18 tok/s
Apple M1 Pro, 16 GB16 GB unified22.4 GiBNo—
Apple M1, 16 GB16 GB unified22.4 GiBNo—
Apple M2 Pro, 16 GB16 GB unified22.4 GiBNo—
Apple M2, 16 GB16 GB unified22.4 GiBNo—
Apple M3, 16 GB16 GB unified22.4 GiBNo—
Apple M4, 16 GB16 GB unified22.4 GiBNo—
GeForce RTX 4060 Ti 16GB16 GB GDDR22.4 GiBOffload19–27 tok/s
GeForce RTX 4070 Ti Super16 GB GDDR22.4 GiBOffload21–31 tok/s
GeForce RTX 408016 GB GDDR22.4 GiBOffload21–31 tok/s
GeForce RTX 4080 Super16 GB GDDR22.4 GiBOffload21–31 tok/s
GeForce RTX 5060 Ti 16GB16 GB GDDR22.4 GiBOffload35–51 tok/s
GeForce RTX 5070 Ti16 GB GDDR22.4 GiBOffload40–59 tok/s
GeForce RTX 508016 GB GDDR22.4 GiBOffload41–59 tok/s
Intel Arc A770 16GB16 GB GDDR622.4 GiBOffload20–30 tok/s
Legion Go16 GB LPDDR5X-750022.4 GiBNo—
Radeon RX 6800 XT16 GB GDDR622.4 GiBOffload20–30 tok/s
Radeon RX 7800 XT16 GB GDDR622.4 GiBOffload21–31 tok/s
Radeon RX 9060 XT 16GB16 GB GDDR622.4 GiBOffload32–46 tok/s
Radeon RX 9070 XT16 GB GDDR622.4 GiBOffload38–55 tok/s
ROG Ally16 GB LPDDR5-640022.4 GiBNo—
Steam Deck LCD16 GB LPDDR5-550022.4 GiBNo—
Steam Deck OLED16 GB LPDDR5-640022.4 GiBNo—
Apple M3 Pro, 18 GB18 GB unified22.4 GiBNo—
Radeon RX 7900 XT20 GB GDDR622.4 GiBOffload63–92 tok/s
Apple M2, 24 GB24 GB unified22.4 GiBNo—
Apple M3, 24 GB24 GB unified22.4 GiBNo—
Apple M4 Pro, 24 GB24 GB unified22.4 GiBNo—
Apple M4, 24 GB24 GB unified22.4 GiBNo—
GeForce RTX 309024 GB GDDR22.4 GiBTight216–349 tok/s
GeForce RTX 409024 GB GDDR22.4 GiBTight233–376 tok/s
Radeon RX 7900 XTX24 GB GDDR622.4 GiBTight222–358 tok/s
ROG Ally X24 GB LPDDR5X-750022.4 GiBNo—
Apple M1 Max, 32 GB32 GB unified22.4 GiBTight92–149 tok/s
Apple M1 Pro, 32 GB32 GB unified22.4 GiBTight46–75 tok/s
Apple M2 Max, 32 GB32 GB unified22.4 GiBTight92–149 tok/s
Apple M2 Pro, 32 GB32 GB unified22.4 GiBTight46–75 tok/s
Apple M4, 32 GB32 GB unified22.4 GiBTight28–45 tok/s
GeForce RTX 509032 GB GDDR22.4 GiBFits414–668 tok/s
Apple M3 Pro, 36 GB36 GB unified22.4 GiBFits35–56 tok/s
Apple M4 Max 14-core, 36 GB36 GB unified22.4 GiBFits95–153 tok/s
Apple M3 Max 16-core, 48 GB48 GB unified22.4 GiBFits92–149 tok/s
Apple M4 Max 16-core, 48 GB48 GB unified22.4 GiBFits126–204 tok/s
Apple M4 Pro, 48 GB48 GB unified22.4 GiBFits63–102 tok/s
Apple M1 Max, 64 GB64 GB unified22.4 GiBFits92–149 tok/s
Apple M2 Max, 64 GB64 GB unified22.4 GiBFits92–149 tok/s
Apple M3 Max 16-core, 64 GB64 GB unified22.4 GiBFits92–149 tok/s
Apple M4 Max 16-core, 64 GB64 GB unified22.4 GiBFits126–204 tok/s
Apple M4 Pro, 64 GB64 GB unified22.4 GiBFits63–102 tok/s
Apple M2 Max, 96 GB96 GB unified22.4 GiBFits92–149 tok/s
Apple M3 Max 16-core, 128 GB128 GB unified22.4 GiBFits92–149 tok/s
Apple M4 Max 16-core, 128 GB128 GB unified22.4 GiBFits126–204 tok/s

Exact answers

GeForce RTX 4090tightGeForce RTX 5090fitsRadeon RX 7900 XTXtightApple M4 Pro, 48 GBfitsApple M1, 8 GBno

FAQ

Why UD-Q4_K_M?

UD-Q4_K_M is the quant this page recommends. The file is Qwen3.6-35B-A3B-UD-Q4_K_M.gguf at 20.6 GiB. Other single-file quants are in the table when a real size was available.

Where do the architecture numbers come from?

text_config num_experts 256, num_experts_per_tok 8, shared_expert_intermediate_size 512, moe_intermediate_size 512. Source: https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/main/config.json

Does this include a long context?

The fit uses 8,192 tokens, or less when that is all the fast memory can hold. The model file allows up to 262,144 tokens, and KV cache grows as you use more of that.

Why is there no Ollama command?

No Ollama library tag was verified for this exact model revision. A nearby tag can be a different checkpoint, so it is left off.

Does this include image input?

No. Weight memory is the cited language-model GGUF. A separate vision projector, when one exists, needs more memory.