Yes — with 4.7 GB to spare
Ornith 1.5 35B-A3B at Q4_K_M fits your CPU only · DDR4 dual-channel entirely on the GPU at 8K context, at an estimated 5.1 tokens per second. Past 246K the KV cache pushes it over — quantise the cache to q8_0, or step down a quantisation, to go longer.
A reasoning-first MIT build on the Qwen3.6 35B-A3B architecture (thinks before every answer). Same VRAM as its base.
The VRAM budget
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 35.5 GB | 36.3 GB | — | ~2.9 | −0.1% ppl | 10.7 GB over |
| Q6_K | 27.4 GB | 28.2 GB | — | ~3.8 | −0.4% ppl | 2.6 GB over |
| Q5_K_M | 23.7 GB | 24.5 GB | 66K | 4.4 | −0.8% ppl | Long context |
| Q4_K_M | 20.2 GB | 20.9 GB | 246K | 5.1 | −1.9% ppl | Recommended |
| Q3_K_M | 16.3 GB | 17.1 GB | 256K | 6.3 | −5.4% ppl | Long context |
| Q2_K | 14.0 GB | 14.8 GB | 256K | 7.4 | −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 10 of its 40 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 ornith-ai/Ornith-1.5-35B-A3B:Q4_K_M \
-c 8192 -ngl 99
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.