Phi-3.5-MoE-instruct vs Qwen-72B
Phi-3.5-MoE-instruct (41.9B parameters) and Qwen-72B (72.3B parameters) side by side: the memory each needs at every precision, what it costs to run on a live GPU, and the context window, KV cache and license where they are published. Numbers are computed from the models' published specs; this page does not rank quality.
Side by side
| Fact | Phi-3.5-MoE-instruct | Qwen-72B |
|---|---|---|
| Parameters | 41.9B | 72.3B |
| Architecture | mixture of 16 experts, 2 active per token | Multi-head attention |
| Context length | 131,072 tokens | 32,768 tokens |
| License | – | – |
| Published precision | BF16 | BF16 |
| VRAM needed, As published | 93.6 GB | 162 GB |
| VRAM needed, FP8 | 46.8 GB | 80.8 GB |
| VRAM needed, INT4 | 23.4 GB | 40.4 GB |
| Cheapest live fit, As published | RTX PRO 6000 · $1.38/hr | RTX A5000 × 7 · $1.23/hr |
| Cheapest live fit, FP8 | L40 · $0.759/hr | RTX PRO 6000 · $1.38/hr |
| Cheapest live fit, INT4 | RTX A5000 · $0.176/hr | RTX A6000 · $0.363/hr |
| KV cache per token (16-bit) | Not published for this architecture | 2.50 MB |
| KV cache at 32k tokens | Not published for this architecture | 80.0 GB |
| KV cache at 128k tokens | Not published for this architecture | 320 GB |
VRAM is the weight size at each precision times a flat 1.2 overhead; see the methodology. The FP8 and INT4 rows need a quantized checkpoint or an engine that quantizes on load. The fit is the lowest-priced single GPU type that holds the model at that precision, or the lowest-priced multi-GPU set (up to 8) when none does. KV cache is for one sequence at 16-bit, computed from each model's config where the attention layout is known.
Which to pick
- Phi-3.5-MoE-instruct needs less VRAM at its published precision (93.6 GB against 162 GB), so it fits on a smaller GPU.
- Phi-3.5-MoE-instruct lists the longer context window (131,072 tokens against 32,768).
- Qwen-72B has the cheaper live GPU fit at its published precision ($1.23/hr against $1.38/hr).
These follow only from the facts in the table above. Whether either model does your task well is a separate question this page does not answer.
Keep reading
- Phi-3.5-MoE-instruct: full VRAM table and live GPU fit
- Qwen-72B: full VRAM table and live GPU fit
- The Phi model series
- The Qwen model series
- All models that fit in 96 GB
- All models that fit in 192 GB
Other comparisons
