Only with CPU offload
Ornith 1.5 35B-A3B at Q4_K_M needs 20.9 GB but only 8.8 GB is addressable, so about 60% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 12 tokens per second — usable for batch work, painful for chat.
A reasoning-first MIT build on the Qwen3.6 35B-A3B architecture (thinks before every answer). Same VRAM as its base.
What hardware do I need for Ornith 1.5 35B-A3B? →
Fits instead: Ornith 1.5 9B (6.1 GB)
The VRAM budget
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 35.5 GB | 36.3 GB | — | ~5.4 | −0.1% ppl | 27.5 GB over |
| Q6_K | 27.4 GB | 28.2 GB | — | ~7.6 | −0.4% ppl | 19.4 GB over |
| Q5_K_M | 23.7 GB | 24.5 GB | — | ~9.3 | −0.8% ppl | 15.7 GB over |
| Q4_K_M | 20.2 GB | 20.9 GB | — | ~12 | −1.9% ppl | 12.1 GB over |
| Q3_K_M | 16.3 GB | 17.1 GB | — | ~17 | −5.4% ppl | 8.3 GB over |
| Q2_K | 14.0 GB | 14.8 GB | — | ~23 | −15% ppl | 6.0 GB over |
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 15
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.