Yes — with 20.5 GB to spare
Ling 3.0 Tiny 7.9B-A1.3B at Q4_K_M fits your CPU only · DDR4 dual-channel entirely on the GPU at 8K context, at an estimated 13 tokens per second. There is room for its full 128K window.
An 8B MoE with 1.3B active and a latent KV cache on only 6 of 24 layers — reasoning and tool use sized for Apple Silicon and edge boxes.
The VRAM budget
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| F16 | 14.7 GB | 15.4 GB | 128K | 3.9 | Reference | Long context |
| Q8_0 | 7.8 GB | 8.5 GB | 128K | 7.4 | −0.1% ppl | Long context |
| Q6_K | 6.0 GB | 6.7 GB | 128K | 9.6 | −0.4% ppl | Long context |
| Q5_K_M | 5.2 GB | 5.9 GB | 128K | 11 | −0.8% ppl | Long context |
| Q4_K_M | 4.4 GB | 5.1 GB | 128K | 13 | −1.9% ppl | Recommended |
| Q3_K_M | 3.6 GB | 4.2 GB | 128K | 16 | −5.4% ppl | Long context |
| Q2_K | 3.1 GB | 3.7 GB | 128K | 19 | −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. This model uses multi-head latent attention, so its cache is a compressed latent rather than full K and V.
How to run it
$ llama-server \
-hf inclusionAI/Ling-3.0-tiny:Q4_K_M \
-c 8192 -ngl 99
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.