Qwen3.5 35B A3B Base — Hardware Requirements & GPU Compatibility
Vision
Qwen3.5-35B-A3B-Base is the mixture-of-experts member of Alibaba's Qwen3.5 base family, pairing a 256-expert MoE layer (8 routed plus 1 shared expert per token) with the same Gated DeltaNet and gated-attention hybrid backbone used across the line. It totals roughly 36 billion parameters but activates only about 3 billion per token (the "A3B" in its name), so decoding stays fast even though the full expert set must stay resident in memory. Like its dense siblings it is pretrain-only, for fine-tuning and research, not direct chat; unlike them, its card lists only a pretraining stage, with no post-training pass. It needs a high-end consumer GPU once quantized. Context length is 262,144 tokens natively, extensible to 1,010,000 tokens. It is released under the Apache 2.0 license, permitting unrestricted commercial and research use, and was published in February 2026, alongside four smaller dense Qwen3.5-Base models from 0.8B to 9B parameters.
Runs well on a 24 GB GPU
Runs from 12 GB (IQ2_S · 32K)Best on 24 GB (Q4_K_M · ~151 tok/s · 128K)Macs from 16 GB (IQ2_S · ~53 tok/s)
Apache 2.0
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run Qwen3.5 35B A3B Base
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
ollama run hf.co/mradermacher/Qwen3.5-35B-A3B-Base-GGUF:Q4_K_MQ4_K_M fits fully on the RTX 4090 · 2.0 GB free · 16K context
Download: Qwen3.5-35B-A3B-Base.Q4_K_M.gguf21.17 GBmradermacher
Ways to runOther ways
each opens step-by-stepRun in cloud
Your own private copy · Q4_K_M · 24 GB GPU
- The easiest way to try this model
- Running in a few minutes, easy setup
- Chat right here or connect any app
$5+ bonus for new users
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 Base 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 | |
| 16.1 GB | 16.4 GB | 3-bit small — 3-bit small quantization | Fits: 256K context · ~184 tok/s | |
| 17.9 GB | 18.2 GB | 3-bit medium — 3-bit medium quantization | Fits: 256K context · ~172 tok/s | |
| 18.4 GB | 18.6 GB | 4-bit · legacy — 4-bit legacy quantization | Fits: 256K context · ~170 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 |
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 Base?
Q4_K_M · 21.9 GBQwen3.5 35B A3B Base (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 Base, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Studio M5 Max (36 GB).
20 GB
Q4_K_M · best fit IQ4_XS (19.8 GB)
1 GPU
RX 7900 XT~75Simulate 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
Alternatives to Qwen3.5 35B A3B Base
Qwen 3.5 family · 34 models →Similar models
6| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| Sage A1.5InternScience | — | 22 GB | ~151 | same size |
| MiniCPM V 4.7 35B A3BOpenBMB | — | 22.1 GB | ~169 | same size |
| Ornith 1.5 35B A3Bornith-ai | — | 22 GB | ~151 | same size |
| Agents A1InternScience | — | 21.5 GB | ~171 | same size |
| Intern S2 PreviewInternLM | — | 73.2 GB | —needs > 24 GB | same size |
| EXAONE 4.5 33BLGAI-EXAONE | — | 23.9 GB | ~33 | same size |
Fine-tunes & variants(2)
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 Base need?
Qwen3.5 35B A3B Base 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 Base?
Yes, at Q5_0 (22.9 GB) or lower. Higher quantizations like Q5_1 (25.1 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.
What's the best quantization for Qwen3.5 35B A3B Base?
For Qwen3.5 35B A3B Base, Q4_K_M (21.9 GB) offers the best balance of quality and VRAM usage. Q5_0 (22.9 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 10.3 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run Qwen3.5 35B A3B Base on a Mac?
Yes — Mac Mini M2 Pro (16 GB) and 133 other Macs can run Qwen3.5 35B A3B Base, the smallest at IQ2_S. 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 Base locally?
Yes — Qwen3.5 35B A3B Base can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is Qwen3.5 35B A3B Base?
At Q4_K_M, Qwen3.5 35B A3B Base 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 Base?
At Q4_K_M, the download is about 21.57 GB. The full-precision BF16 version is 71.90 GB. The smallest option (IQ2_XXS) is 9.89 GB.
Which GPUs can run Qwen3.5 35B A3B Base?
7 consumer GPUs can run Qwen3.5 35B A3B Base at Q4_K_M. Top options include AMD Radeon RX 7900 XTX, NVIDIA GeForce RTX 3090.
Which devices can run Qwen3.5 35B A3B Base?
100 devices with unified memory can run Qwen3.5 35B A3B Base 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.