Only with CPU offload
Ornith 1.5 397B-A17B at Q4_K_M needs 224.1 GB but only 38.4 GB is addressable, so about 83% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 1.7 tokens per second — usable for batch work, painful for chat.
The flagship Ornith on the Qwen3.5 397B-A17B architecture, MIT-licensed. A 256 GB Mac Studio at Q4, and it is in the Ollama library.
What hardware do I need for Ornith 1.5 397B-A17B? →
Fits instead: Ornith 1.5 35B-A3B (20.9 GB) · Ornith 1.5 9B (6.1 GB)
The VRAM budget
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 392.8 GB | 393.7 GB | — | ~0.9 | −0.1% ppl | 355.3 GB over |
| Q6_K | 303.2 GB | 304.0 GB | — | ~1.2 | −0.4% ppl | 265.6 GB over |
| Q5_K_M | 262.0 GB | 262.9 GB | — | ~1.4 | −0.8% ppl | 224.5 GB over |
| Q4_K_M | 223.2 GB | 224.1 GB | — | ~1.7 | −1.9% ppl | 185.7 GB over |
| Q3_K_M | 180.7 GB | 181.5 GB | — | ~2.3 | −5.4% ppl | 143.1 GB over |
| Q2_K | 154.8 GB | 155.7 GB | — | ~2.7 | −15% ppl | 117.3 GB over |
Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it. Only 15 of its 60 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-1.5-397B:Q4_K_M \
-c 8192 -ngl 10
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.