Qwen3.5 35B A3B — Hardware Requirements & GPU Compatibility
Vision
Qwen3.5 35B A3B is Alibaba's mixture-of-experts vision-language model in the Qwen3.5 line, totaling 36 billion parameters with about 3 billion active per token (the A3B in its name). It accepts text, images, and video as input, using a hybrid linear-attention and sparse-MoE architecture built for efficiency. Only the active parameters compute per token, so generation is faster than a same-sized dense model, though all 36 billion weights must fit in memory, calling for a high-end 24-32GB-class GPU once quantized. The model offers a 262K token context window for long documents or multi-turn chats. It is released under the Apache 2.0 license, permitting unrestricted commercial and research use, and was published in February 2026 as part of Alibaba's push toward efficient multimodal MoE models.
Runs well on a 24 GB GPU
Runs from 16 GB (Q2_K · 32K)Best on 24 GB (Q4_K_M · ~151 tok/s · 128K)Macs from 24 GB (Q3_K_M · ~53 tok/s)
Score 105.8#44 of 106 open#6 of 36 that fit 24 GBApache 2.0
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
llmrun Composite Score →
105.8
- 90% range 101.6–110.1
- #44 of 106 open models
- #6 of 36 that fit 24 GB
Run Qwen3.5 35B A3B
One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.
Recommended
vllm serve Qwen/Qwen3.5-35B-A3B --max-model-len 16384vLLM repo with its own quantisation
Download: Weights on Hugging Face
Ways to runOther ways
each opens step-by-stepRun in cloud
Runs on RTX PRO 4500 (32 GB) (10 GB headroom) · Q4_K_M
- Generation speed
- ~139 tok/s
- generation speed
- Cost per 1M output tokens
- $0.68
- per 1M output tokens
How Much VRAM Does Qwen3.5 35B A3B Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 15.7 GB | 15.9 GB | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Fits: 256K context · ~187 tok/s | |
| 17.9 GB | 18.2 GB | 3-bit medium — 3-bit medium quantization | Fits: 256K context · ~172 tok/s | |
| 21.9 GB | 22.2 GB | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Fits: 128K context · ~151 tok/s | |
| 26 GB | 26.3 GB | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Does not fit: needs 26.1 GB | |
| 30.0 GB | 30.3 GB | 6-bit, very good — 6-bit quantization, very good quality | Does not fit: needs 30.1 GB | |
| 36.3 GB | 36.6 GB | 8-bit, near-lossless — 8-bit quantization, near-lossless | Does not fit: needs 36.4 GB | |
| 72.3 GB | 72.5 GB | 16-bit · bfloat — Brain floating point 16 — preferred for training | Does not fit: needs 72.4 GB |
Selected Q4_K_M
4-bit medium quantization22.0GBat 16K context
- Weights
- 21.6 GB
- KV cache
- 0.2 GB
- Overhead
- 0.3 GB
Fits RTX 4090 (reference) · 2.0 GB free · ~151 tok/s
≥50 tok/s needs ≥250 GB/s · RTX 5090 ~213, RTX 3090 Ti ~151
Or .
Which Hardware Can Run Qwen3.5 35B A3B?
Q4_K_M · 21.9 GBQwen3.5 35B A3B (Q4_K_M) requires 21.9 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 29+ GB is recommended. Using the full 262K context window can add up to 2.6 GB, bringing total usage to 24.6 GB. 7 GPUs can run it, including NVIDIA GeForce RTX 5090. 100 devices with unified memory can run Qwen3.5 35B A3B, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Studio M5 Max (36 GB).
20 GB
Q4_K_M · best fit Q3_K_M (18.0 GB)
1 GPU
RX 7900 XT~77Simulate output on RX 7900 XT32–40 GB
fits · 10.0 GB free · up to 256K ctx
1 GPU · 3 Mac classes · 4 device families
RTX 5090~213Simulate output on RTX 5090M5 Max36 GB~59Simulate output on Mac Studio M5 Max (36 GB)M532 GB~29Simulate output on MacBook Air 13" M5 (32 GB)Orin~42Simulate output on NVIDIA Jetson AGX Orin 32GBSnapdragon X Elite (X1E-80-100)~27Simulate output on Snapdragon X Elite Copilot+ PC
By VRAM tier · typical GPUs and Macs at Q4_K_M
How good is Qwen3.5 35B A3B?
The llmrun Score merges 7 public benchmark boards into a single figure, where 100 represents the average for open models.
Score vs VRAM in the 24 GB tier
- Qwen3.5 35B A3B
- Other models
- Best Score for its VRAM
#6 of 36 that fit 24 GB
Each dot is a model at the best quant that fits a 24 GB GPU. The line steps up to the best Score available at each amount of VRAM; a dot below it is beaten by a model that needs no more memory.
Full 24 GB ranking →Benchmarks
The 6 boards that weigh most in the llmrun Score, of 9 with a result for this model; badges show who ran the test.
- GPQA Diamond
Epoch AI · Feb 2026Independent
Independent83.5%
#25/100 open
- LMArena Text
LMArena · Oct 2026Arena
Arena1393.8
#55/188 open
- AIME 2024/2025
Epoch AI · Feb 2026Independent
Independent70.0%
#31/88 open
- LMArena WebDev
LMArena · Oct 2026Arena
Arena1254.2
#51/55 open
- MMLU-Pro
TIGER-LabIndependent
Independent85.3%
#7/146 open
- SciCode
Epoch AI · Feb 2026Lab-reported
Lab-reported29.3%
#60/72 open
In the 24 GB class
36 models fit 24 GBSame Score, only models that fit 24 GB at their best quant: the ranking a 24 GB owner actually chooses from.
- #1
Qwen3.8 27BAlibaba
122.0
~38 tok/s · 17.9 GB
- #2
Qwen3.6 27BAlibaba
111.2
~38 tok/s · 17.9 GB
- #3
Qwen3.6 35B A3BAlibaba
106.8
~151 tok/s · 22.0 GB
- #4
Gemma 4 31B ITGoogle
106.7
~32 tok/s · 21.7 GB
- #5
Gemma 4 26B A4B ITGoogle
105.8
~171 tok/s · 16.5 GB
Alternatives to Qwen3.5 35B A3B
Qwen 3.5 family · 34 models →Official models with an llmrun Score, each answering one question; numbers are at the best quant that fits a 24 GB GPU.
Best alternative at this size
Qwen3.8 27B
Alibaba
- 16.2 higher Score
- 4.2 GB less VRAM
- 113 tok/s slower
Score 122.0 · 17.9 GB · ~38 tok/s
Fits in less VRAM
GPT OSS 20B
OpenAI
- 9.4 lower Score
- 8.9 GB less VRAM
- 39 tok/s faster
Score 96.4 · 13.1 GB · ~190 tok/s
Other vendor
Gemma 4 31B IT
Google
- 0.9 higher Score
- 0.3 GB less VRAM
- 119 tok/s slower
Score 106.7 · 21.7 GB · ~32 tok/s
Similar models
10| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| Qwen3.6 27BAlibaba | 111.2 | 17.9 GB | ~38 | similar score |
| Qwen3.6 35B A3BAlibaba | 106.8 | 22 GB | ~151 | similar score |
| Gemma 4 26B A4B ITGoogle | 105.8 | 16.5 GB | ~171 | similar score |
| Qwen3.5 27BAlibaba | 103.8 | 17.9 GB | ~38 | similar score |
| Seed OSS 36B InstructByteDance-Seed | 102.5 | 22.8 GB | ~31 | similar score |
| Qwen3 30B A3B Thinking 2507Alibaba | 100.7 | 19.4 GB | ~173 | similar score |
| Qwen3 14BAlibaba | 98.0 | 11.9 GB | ~69 | similar score |
| Qwen3 8BAlibaba | 93.9 | 7.6 GB | ~119 | similar score |
| Qwen3 32BAlibaba | 98.6 | 22.6 GB | ~32 | same size |
| DeepSeek R1 Distill Qwen 32BDeepSeek | 97.2 | 23 GB | ~34 | same size |
Fine-tunes & variants(1)
Specifications
- Publisher
- Alibaba
- Family
- Qwen 3.5
- Parameters
- 36.0B
- Active parameters
- 4.7B per token
- Architecture
- Qwen3_5MoeForConditionalGeneration
- Context Length
- 262,144 tokens
- Vocabulary Size
- 248,320
- Release Date
- 2026-02-24
- License
- Apache 2.0
Frequently Asked Questions
How much VRAM does Qwen3.5 35B A3B need?
Qwen3.5 35B A3B requires 21.9 GB of VRAM at Q4_K_M, or 72.3 GB at BF16. Full 262K context adds up to 2.6 GB (24.6 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 36.0B × 4.8 bits ÷ 8 = 21.6 GB
KV Cache + Overhead ≈ 0.3 GB (at 2K context + ~0.3 GB framework)
Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.
KV Cache + Overhead ≈ 2.9 GB (at full 262K context)
VRAM usage by quantization
Can NVIDIA GeForce RTX 4090 run Qwen3.5 35B A3B?
Yes, at Q4_K_M (21.9 GB) or lower. Higher quantizations like Q5_K_M (26 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.
What's the best quantization for Qwen3.5 35B A3B?
For Qwen3.5 35B A3B, Q4_K_M (21.9 GB) offers the best balance of quality and VRAM usage. Q5_K_M (26 GB) provides better quality if you have the VRAM. The smallest option is Q2_K at 15.7 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run Qwen3.5 35B A3B on a Mac?
Yes — Mac Mini M5 Pro (24 GB) and 106 other Macs can run Qwen3.5 35B A3B, the smallest at Q3_K_M. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 21.9 GB of usable unified memory (RAM minus macOS overhead), from Mac Studio M1 Max (32 GB).
Can I run Qwen3.5 35B A3B locally?
Yes — Qwen3.5 35B A3B can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is Qwen3.5 35B A3B?
At Q4_K_M, Qwen3.5 35B A3B can reach ~116 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~151 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = 1000 ÷ (active GB ÷ (bandwidth GB/s × efficiency) × 1000 + layers × routing ms)
Mixture-of-Experts: just the active experts get read for each token, along with a fixed routing cost per layer.
Example: NVIDIA B200 → 2.9 GB active ÷ (8000 × 0.65) = 0.557 ms, plus 40 layers × 0.055 ms = 2.200 ms, so 1000 ÷ 2.757 ms = ~363 tok/s
Estimated speed at Q4_K_M (21.9 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of Qwen3.5 35B A3B?
At Q4_K_M, the download is about 21.57 GB. The full-precision BF16 version is 71.90 GB. The smallest option (Q2_K) is 15.28 GB.
Which GPUs can run Qwen3.5 35B A3B?
7 consumer GPUs can run Qwen3.5 35B A3B at Q4_K_M. Top options include AMD Radeon RX 7900 XTX, NVIDIA GeForce RTX 3090.
Which devices can run Qwen3.5 35B A3B?
100 devices with unified memory can run Qwen3.5 35B A3B at Q4_K_M (21.9 GB), including AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.