Yes — with 23.6 GB to spare
Qwen3.5 2B at Q4_K_M fits your CPU only · DDR4 dual-channel entirely on the GPU at 8K context, at an estimated 17 tokens per second. There is room for its full 256K window.
Fully on GPU
8K context
Q4_K_M · 1.3 GB
Apache 2.0
Released 28 Feb 2026
Vision
Phone-class, and multimodal. Replaces Llama 3.2 3B as the "it runs on anything" answer.
The VRAM budget
weights 1.3 GB
Weights 1.3 GB
KV cache @ 8K 0.09 GB
Runtime overhead 0.6 GB
Free 23.6 GB of 25.6 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| F16 | 4.2 GB | 4.9 GB | 256K | 5.1 | Reference | Long context |
| Q8_0 | 2.2 GB | 2.9 GB | 256K | 9.5 | −0.1% ppl | Long context |
| Q6_K | 1.7 GB | 2.4 GB | 256K | 12 | −0.4% ppl | Long context |
| Q5_K_M | 1.5 GB | 2.2 GB | 256K | 14 | −0.8% ppl | Long context |
| Q4_K_M | 1.3 GB | 2.0 GB | 256K | 17 | −1.9% ppl | Recommended |
| Q3_K_M | 1.0 GB | 1.7 GB | 256K | 21 | −5.4% ppl | Long context |
| Q2_K | 0.9 GB | 1.6 GB | 256K | 24 | −15% ppl | Long context |
Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it. Only 6 of its 24 blocks keep a per-token KV cache; the rest are linear-attention, Mamba or convolution blocks with a fixed-size state.
How to run it
$ llama-server \
-hf Qwen/Qwen3.5-2B:Q4_K_M \
-c 8192 -ngl 99
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.
01Download is 1.3 GB. Keep it on an SSD — a first load off a spinning disk takes minutes.
02With no GPU, thread count matters more than clock. Start at one thread per physical core.
03There is room to go to the model's full 256K context on this card.