Qwen 3.5 72B Instruct
Apache-2.0 licensed 72.7B dense powerhouse for multilingual engineering, complex SQL/data synthesis, and local dual-GPU workstation deployment.
Provider
Alibaba Qwen (Apache 2.0)
Architecture
Dense Transformer (72.7B) + GQA (8 KV Heads) + YaRN
Parameters
72.7B Dense
Context Window
131,072 tokens (Max Output: 32,768 tokens)
API Pricing
Input: $0.35 / 1M · Cached: $0.15 / 1M · Output: $0.90 / 1M
VRAM (16K Ctx)
Q4_K_M: 47.7 GB · Q8_0: 81.4 GB · FP16: 152.04 GB
Hardware Target
Dual RTX 4090/3090 (48GB) or Apple M4 Max (64GB)
Interactive Token Cost Estimator (Qwen 3.5 72B)
GPU Hardware Compatibility Matrix (Qwen 3.5 72B at 16K Context)
| GPU / Hardware Profile | VRAM | Q4_K_M (4.8-bit) | Q8_0 (8-bit) | Est. Speed (Q4) |
|---|---|---|---|---|
| RTX 5060 Ti 16GB 448 GB/s | 16 GB | REQUIRES MULTI-GPU (47.7 GB) | REQUIRES MULTI-GPU (81.4 GB) | ~1 tok/s |
| RTX 4090 24GB 1008 GB/s | 24 GB | REQUIRES MULTI-GPU (47.7 GB) | REQUIRES MULTI-GPU (81.4 GB) | ~3 tok/s |
| RTX 5090 32GB 1792 GB/s | 32 GB | REQUIRES MULTI-GPU (47.7 GB) | REQUIRES MULTI-GPU (81.4 GB) | ~6 tok/s |
| M4 Max 64GB 546 GB/s | 52 GB | TIGHT FIT (Q8 KV) (47.7 GB) | REQUIRES MULTI-GPU (81.4 GB) | ~6 tok/s |
| 2× 24GB (48GB) 920 GB/s | 48 GB | TIGHT FIT (Q8 KV) (47.7 GB) | REQUIRES MULTI-GPU (81.4 GB) | ~10 tok/s |
| H100 80GB 3350 GB/s | 80 GB | FITS IN VRAM (47.7 GB) | TIGHT FIT (Q8 KV) (81.4 GB) | ~50 tok/s |
| M4 Ultra 192GB 819 GB/s | 164 GB | FITS IN VRAM (47.7 GB) | FITS IN VRAM (81.4 GB) | ~12 tok/s |
| 8× H200 (1,128GB) 4800 GB/s | 1128 GB | FITS IN VRAM (47.7 GB) | FITS IN VRAM (81.4 GB) | ~72 tok/s |
Architectural Strengths & Target Workloads
- Fits comfortably in 48GB VRAM (Dual RTX 3090/4090) or 64GB Apple M4 Max at Q4_K_M (~44 GB weights)
- Grouped-Query Attention (GQA) keeps KV-cache predictable up to 64K–128K context
- Permissive Apache 2.0 license for commercial self-hosting
Primary Engineering Workloads: Self-Hosted Workstation Assistant · Complex SQL & System Synthesis · Multilingual Technical Documentation.
Frequently Asked Questions (Qwen 3.5 72B)
How much does Qwen 3.5 72B Instruct cost per 1M tokens?
Qwen 3.5 72B Instruct is priced at $0.35 per 1M uncached input tokens, $0.15 per 1M cached input tokens, and $0.90 per 1M output tokens (OpenRouter / Together / Self-Hosted Dual-GPU or Apple M4 Max 64GB).
What is the context window of Qwen 3.5 72B Instruct?
It supports up to 131,072 tokens (131,072 tokens) and a maximum generation output of 32,768 tokens.