Alibaba
Qwen3.5 9B
9,7B
Parameters
5,7 Go
Q4_K_M file
≥ 7,2 Go
Recommended VRAM
262 144
Max context
Speed by graphics card
Simulate in the “Configurator” →Standard llama-bench run (prompt 512 / generation 128 tokens), model fully in VRAM. Click a card for details. Electricity cost at the default price (0,25 €/kWh).
| Card · Q4_K_M | VRAM | Generation | Prompt | VRAM used | Power | € elec. / M tok | Price |
|---|---|---|---|---|---|---|---|
| RTX 5070 Ti | 16 Go | 125 tok/s | 5 396 | 7,1 Go | 152 W | 0,08 €/M tok | — |
Not measured on these cards yet — click for an estimate:
Technical sheet
- Published
- 2026-02-27
- Architecture
- 32 layers · 4 KV heads · dimension 256
- Attention
- Hybrid: 8 of 32 layers use standard attention, the others a fixed-memory mechanism (Mamba, linear attention) — a very light cache.
- Context cache (f16)
- ≈ 31 MB per 1,000 tokens · 1,0 Go for 32,768 tokens
- GGUF file
- bartowski/Qwen_Qwen3.5-9B-GGUF · Qwen_Qwen3.5-9B-Q4_K_M.gguf Download (5,7 Go)
- Official repo
- Qwen/Qwen3.5-9B
Public Hugging Face data (official config.json, GGUF repo). The context cache only counts full-attention layers.