Only with CPU offload
Ornith 1.5 35B-A3B at Q4_K_M needs 20.9 GB but only 18.4 GB is addressable, so about 13% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 28 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.
Hardware requirements What hardware do I need to run Ornith 1.5 35B-A3B? →At Q4_K_M with an 8K context: a 32 GB GPU (RTX 5090) or a 36 GB Mac (M4 Max). Minimum, recommended, high-quality and long-context tiers, with all 114 devices checked at every quantisation.
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 | — | ~7.3 | −0.1% ppl | 17.9 GB over |
| Q6_K | 27.4 GB | 28.2 GB | — | ~12 | −0.4% ppl | 9.8 GB over |
| Q5_K_M | 23.7 GB | 24.5 GB | — | ~17 | −0.8% ppl | 6.1 GB over |
| Q4_K_M | 20.2 GB | 20.9 GB | — | ~28 | −1.9% ppl | 2.5 GB over |
| Q3_K_M | 16.3 GB | 17.1 GB | 74K | 56 | −5.4% ppl | Long context |
| Q2_K | 14.0 GB | 14.8 GB | 194K | 65 | −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 34
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.
Related models on an RTX 4000 Ada
| Model | Total | Tok/s est. | |
|---|---|---|---|
| Ornith 1.5 9B 9.41B · same family |
Q4_K_M · 6.1 GB | 41 | Runs great |
| Ornith 1.5 397B-A17B 397B MoE · same family |
Q4_K_M · 224.1 GB | — | Won't fit |
| Qwen3.6 35B-A3B 35.9B MoE · similar size |
Q4_K_M · 20.9 GB | ~28 | CPU offload |
| LLM-jp 4 33B Thinking 33.2B · similar size |
Q4_K_M · 21.3 GB | ~6.6 | CPU offload |
| Qwen3 32B 32.8B · similar size |
Q4_K_M · 21.0 GB | ~6.9 | CPU offload |
| DeepSeek-R1-Distill-Qwen 32B 32.8B · similar size |
Q4_K_M · 21.0 GB | ~6.9 | CPU offload |
| Qwen2.5-Coder 32B 32.8B · similar size |
Q4_K_M · 21.0 GB | ~6.9 | CPU offload |
Ornith 1.5 35B-A3B on other hardware
| Device | Total | Tok/s est. | |
|---|---|---|---|
| GeForce RTX 4060 Laptop 7.0 GB usable · 256 GB/s |
20.9 GB | — | Won't fit |
| GeForce RTX 3060 12 GB 10.6 GB usable · 360 GB/s |
20.9 GB | ~13 | CPU offload |
| M4 · 16 GB 10.7 GB usable · 120 GB/s |
20.9 GB | — | Won't fit |
| GeForce RTX 4090 22.4 GB usable · 1008 GB/s |
20.9 GB | 126 | Tight fit |
| Radeon RX 7900 XTX 22.4 GB usable · 960 GB/s |
20.9 GB | 120 | Tight fit |
| GeForce RTX 5090 30.4 GB usable · 1792 GB/s |
20.9 GB | 225 | Runs great |
| M4 Max · 36 GB 27.0 GB usable · 410 GB/s |
20.9 GB | 58 | Runs great |
| M3 Max · 64 GB 48.0 GB usable · 400 GB/s |
20.9 GB | 56 | Runs great |