Qwen 3.5 Hardware Requirements: VRAM for 9B, 27B & 35B [2026]

TL;DR: Qwen 3.5 9B needs ~6.5GB of VRAM at 4-bit — any 8GB GPU works. The 27B needs ~17GB — that's RTX 3090/RX 7900 XT territory. The 35B-A3B MoE needs ~22GB to load but decodes like a 3B model, making it the speed king on a 24GB card. Qwen 2.5 72B (a different, older family) is a 47GB file at Q4_K_M and needs dual GPUs or 64GB+ unified memory.

Memory requirements per Unsloth's Qwen 3.5 documentation:

On this page:

• Qwen 3.5 vs Qwen 2.5: Which Sizes Actually Exist
• VRAM Requirements by Model Size and Quantization
• GPU Recommendations Per Tier
• Tokens Per Second on Common Hardware
• Apple Silicon: MLX vs llama.cpp
• Choosing the Right Tier for Your Hardware

Qwen 3.5 vs Qwen 2.5: Which Sizes Actually Exist

The two families get mixed up constantly in searches, so let's settle it:

• Qwen 3.5 (released February 2026, Apache 2.0, natively multimodal, 262K context): 0.8B, 2B, 4B, 9B, 27B, 35B-A3B, 122B-A10B, and the 397B-A17B flagship. The "A" sizes are sparse Mixture-of-Experts — the 35B-A3B model card lists 35B total parameters with only 3B active per token.
• Qwen 2.5 (the older, text-only family): 0.5B, 1.5B, 3B, 7B, 14B, 32B, and 72B, plus the Coder fine-tunes up to 32B.

There is no Qwen 2.5 9B or 27B, and no Qwen 3.5 72B. This guide covers the four sizes people actually search for: Qwen 3.5's 9B, 27B, and 35B-A3B, plus Qwen 2.5's 72B.

VRAM Requirements by Model Size and Quantization

The Qwen 3.5 figures below are from Unsloth's documentation; they cover model weights at each precision. Add headroom for KV cache — a few GB at modest context, more if you push toward the 262K native window.

Qwen 3.5 9B

Minimum GPU: RTX 3060 8GB for 4-bit. RTX 3060 12GB or better for 6-bit.

The 9B is the "just works" tier — and reviewers consistently flag it as the standout of the family for its size. Any gaming GPU from the last four years handles 4-bit. If your card has 12GB, the 6-bit quant is worth the extra memory.

Qwen 3.5 27B

Minimum GPU: a 20GB+ card for 4-bit — RX 7900 XT 20GB or RTX 3090 24GB. A 16GB card can run the 3-bit quant, tightly.

The dense 27B is the quality pick for users with 20–24GB cards who want consistent (non-MoE) behavior across all workloads.

Qwen 3.5 35B-A3B (Sparse MoE)

Minimum GPU: RTX 3090 24GB (tight at 4-bit) or RTX 4090 24GB.

Here's why this model matters: it loads like a 35B but decodes like a 3B. Only 3B parameters are active per generated token, so the per-token memory read is a fraction of what the dense 27B requires. If your card can hold the weights, the 35B-A3B is both the smarter and the faster model — it's the one Ollama picked to launch its MLX backend on Apple Silicon.

Qwen 2.5 72B

File sizes from the bartowski Qwen2.5-72B-Instruct GGUF repository.

This is where consumer hardware hits its ceiling. Dual RTX 3090/4090 (48GB combined) fits Q4_K_M with almost no headroom — many builders drop to Q4_K_S or IQ4_XS for KV cache room. On a single 24GB card, your option is CPU+GPU hybrid inference (see our llama.cpp hybrid inference guide) at a fraction of the speed. For general VRAM planning, see how much VRAM do you need or the VRAM calculator.

GPU Recommendations Per Tier

For Qwen 3.5 9B

Best value: RTX 3060 12GB — handles the 6-bit quant with headroom for KV cache, commonly listed used around $200–230.

Best performance: RTX 4060 Ti 16GB — runs 8-bit comfortably and future-proofs toward the 27B's 3-bit quant.

AMD option: RX 7600 8GB for 4-bit (tight) or a 12GB card for 6-bit. ROCm support is solid on Linux for these cards.

For Qwen 3.5 27B and 35B-A3B

Sweet spot: RX 7900 XT 20GB — runs the 27B at 4-bit (17GB) with room to spare; better VRAM per dollar than an RTX 3090 if buying new.

Best single-card option: RTX 3090 24GB used or RTX 4090 24GB new — both run the 27B at 4-bit comfortably and the 35B-A3B at 4-bit (22GB) with modest context.

AMD: RX 7900 XTX 24GB — same capacity story as the 3090, with mature ROCm 6.x support on Linux.

For Qwen 2.5 72B

Dual GPU: Dual RTX 3090 (48GB combined) just fits Q4_K_M. llama.cpp's tensor split distributes layers automatically — see our multi-GPU inference guide.

Apple Silicon: 64GB unified memory handles Q4_K_M; 96GB+ handles Q8_0.

AMD Strix Halo (mini PCs): 96GB unified-memory configs handle 72B Q4_K_M fully in GPU-addressable memory once GTT is configured (Linux — see the AMD Ryzen AI Max GTT guide).

Tokens Per Second on Common Hardware

Decode speed on consumer GPUs is limited by memory bandwidth, not compute. That gives you a ceiling you can calculate yourself: bandwidth ÷ bytes read per token. For dense models, bytes per token ≈ the weight file size; real-world results typically land at 60–80% of the ceiling.

The MoE row is the story: because only ~3B parameters activate per token, the 35B-A3B reads a few GB per token instead of 22 — which is how Ollama's MLX backend posts 112 tok/s on Apple Silicon for this exact model.

Apple Silicon: MLX vs llama.cpp

Apple Silicon is uniquely suited to the Qwen families because of unified memory — the same pool serves CPU and GPU with no PCIe transfer. Unsloth notes the 35B-A3B fits on a 22GB-class Mac and the 27B on an 18GB-class device at 4-bit.

For Qwen 3.5 specifically, MLX is the fast path — it's the framework behind Ollama's Apple Silicon preview, which roughly doubled decode speed over the old llama.cpp Metal path (58 → 112 tok/s on M5-generation hardware, per Ollama's announcement). Running MLX directly:

pip install mlx-lm
mlx_lm.generate --model Qwen/Qwen3.5-35B-A3B --prompt "Explain transformer attention"

For the runtime trade-offs, see our MLX vs llama.cpp comparison. For Qwen 2.5 72B: 64GB unified memory handles Q4_K_M, and 96GB+ configurations handle Q8_0.

Choosing the Right Tier for Your Hardware

One more thing: the Qwen 2.5 Coder models (up to 32B) follow the same VRAM requirements as their base counterparts — for that specific build, see Qwen 2.5 Coder 32B hardware requirements. And for grabbing the right quantized file for your card, see the HuggingFace GGUF download guide.