Qwen 3.6 VRAM & RAM requirements — 27B vs 35B-A3B
Estimated memory for every GGUF quant of Qwen 3.6 27B and Qwen 3.6 35B-A3B: what fits fully on a GPU, what long context really costs, and which phones can run them. Total memory = weights × 1.05 overhead + KV cache at 4k context + 0.6GB runtime — the download size alone understates it.
Every number here is an estimate from memory math, not a measured benchmark — see methodology.
- Qwen 3.6 27B: ~20.1GB at Q4_K_M → a 24GB GPU; no quant fits a 16GB card without CPU offload.
- Qwen 3.6 35B-A3B: ~26.3GB at Q4_K_M → a 32GB GPU for full offload, but as a MoE (3B active) it runs on 8GB cards with experts in system RAM.
- Phones: neither fits any of the 99 phones we track — for on-phone Qwen, check what your phone can run.
Qwen 3.6 27B: memory by quant
27B dense parameters. ★ = the size/quality sweet spot we recommend. Full per-phone verdicts on the Qwen 3.6 27B page.
| Quant | Download | Est. memory @4k | Min GPU (full offload) |
|---|---|---|---|
| Q3_K_M | 13.6 GB | ~16.8 GB | 24 GB |
| IQ4_XS | 15.4 GB | ~18.7 GB | 24 GB |
| Q4_0 | 15.8 GB | ~19.1 GB | 24 GB |
| Q4_K_S | 15.9 GB | ~19.2 GB | 24 GB |
| Q4_K_M ★ | 16.8 GB | ~20.1 GB | 24 GB |
| Q5_K_M | 19.5 GB | ~23 GB | 24 GB |
| Q6_K | 22.5 GB | ~26.1 GB | 32 GB |
| Q8_0 | 28.6 GB | ~32.5 GB | 48 GB |
Qwen 3.6 35B-A3B: memory by quant
35B total parameters, 3B active per token (mixture-of-experts). ★ = the size/quality sweet spot we recommend. Full per-phone verdicts on the Qwen 3.6 35B-A3B page.
| Quant | Download | Est. memory @4k | Min GPU (full offload) |
|---|---|---|---|
| Q3_K_M | 16.6 GB | ~20.5 GB | 24 GB |
| IQ4_XS | 17.7 GB | ~21.6 GB | 24 GB |
| Q4_K_S | 20.9 GB | ~25 GB | 32 GB |
| MXFP4 | 21.7 GB | ~25.8 GB | 32 GB |
| Q4_K_M ★ | 22.1 GB | ~26.3 GB | 32 GB |
| Q5_K_M | 26.5 GB | ~30.9 GB | 32 GB |
| Q6_K | 29.3 GB | ~33.8 GB | 48 GB |
| Q8_0 | 36.9 GB | ~41.8 GB | 48 GB |
What fits in your GPU
Largest quant that fits fully on-GPU at 4k context. "Offload" means llama.cpp can still run it by keeping part of the model in system RAM — expect a large speed penalty for the dense 27B, a modest one for the MoE 35B-A3B (its experts stream well).
| GPU VRAM | Qwen 3.6 27B | Qwen 3.6 35B-A3B |
|---|---|---|
| 8 GB | offload only | offload only |
| 12 GB | offload only | offload only |
| 16 GB | offload only | offload only |
| 24 GB | Q5_K_M (~23 GB) | IQ4_XS (~21.6 GB) |
| 32 GB | Q6_K (~26.1 GB) | Q5_K_M (~30.9 GB) |
| 48 GB | Q8_0 (~32.5 GB) | Q8_0 (~41.8 GB) |
The hidden cost: context length
Most quoted "VRAM requirements" assume a short context. The KV cache grows linearly with context length and at 128k it dwarfs the Q4 weights. Estimated KV cache alone:
| Context | 27B KV cache | 27B total (Q4_K_M) | 35B-A3B KV cache | 35B-A3B total (Q4_K_M) |
|---|---|---|---|---|
| 4k | ~1.9 GB | ~20.1 GB | ~2.5 GB | ~26.3 GB |
| 16k | ~7.6 GB | ~25.8 GB | ~9.8 GB | ~33.6 GB |
| 32k | ~15.1 GB | ~33.4 GB | ~19.6 GB | ~43.4 GB |
| 128k | ~60.5 GB | ~78.7 GB | ~78.4 GB | ~102.2 GB |
Estimated from parameter count; actual KV size depends on the model's attention config and quantized-cache settings.
Can a phone run Qwen 3.6?
No. At Q4_K_M the 27B needs ~20.1GB of usable memory and 35B-A3B ~26.3GB — more than any of the 99 phones we track can give a single app. For on-phone Qwen, start from Qwen 3.5 9B and below, or check what your phone can run.
Not sure what those speeds feel like? Try tokens/s yourself.
FAQ
How much VRAM do I need to run Qwen 3.6 27B?
For full GPU offload at 4k context, Qwen 3.6 27B needs ~20.1GB at Q4_K_M — in practice a 24GB card (RTX 3090/4090 class). The smallest quant we track (Q3_K_M, 13.6GB download) still needs ~16.8GB in total, so 16GB cards can only run it with partial CPU offload, at reduced speed. These are estimates from memory math, not measurements.
How much VRAM does Qwen 3.6 35B-A3B need?
~26.3GB at Q4_K_M for full GPU offload at 4k context. But 35B-A3B is a mixture-of-experts model with only 3B active parameters, so llama.cpp can keep the shared layers on an 8GB GPU and stream experts from system RAM — slower than full offload, but usable. That's why you see reports of it running on small cards.
Can I run Qwen 3.6 on 16GB of VRAM?
Not fully on-GPU at 4k context: the smallest 27B quant needs ~16.8GB and the smallest 35B-A3B quant ~20.5GB. On 16GB you either shorten the context, offload part of the model to system RAM (llama.cpp does this automatically), or pick a smaller model like Qwen 3.5 9B.
How much memory does long context add?
The KV cache grows linearly with context. For 27B we estimate ~1.9GB at 4k but ~60.5GB at 128k — more than the Q4 weights themselves. Quoted "VRAM requirements" that ignore context length understate what long-context use actually needs.
Can my phone run Qwen 3.6?
Not the phones we track: at Q4_K_M the 27B needs ~20.1GB of usable memory and 35B-A3B ~26.3GB, beyond current phone RAM. Smaller Qwen models (3.5 9B and below) do run on phones.
Qwen 3.6 27B or 35B-A3B — which should I run?
Memory needs are similar (~20.1GB vs ~26.3GB at Q4, 4k context), but 35B-A3B decodes much faster because only 3B parameters are active per token, and it degrades more gracefully on small GPUs via expert offload. Pick 27B (dense) when you want maximum quality per GB on a machine that fits it fully.