Yes, just — 0.9 GB spare
Ornith 1.5 9B at Q4_K_M fits your Arc A750 entirely on the GPU at 8K context, at an estimated 59 tokens per second. Past 35K the KV cache pushes it over — quantise the cache to q8_0, or step down a quantisation, to go longer.
Fully on GPU
8K context
Q4_K_M · 5.3 GB
MIT
Released 19 Aug 2026
New this week
The small Ornith: a coding-agent reasoning build on the Qwen3.5 9B architecture. Same VRAM as its base, thinks before every answer.
The VRAM budget
weights 5.3 GB
Weights 5.3 GB
KV cache @ 8K 0.25 GB
Runtime overhead 0.6 GB
Free 0.9 GB of 7.0 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| F16 | 17.5 GB | 18.4 GB | — | ~3.0 | Reference | 11.4 GB over |
| Q8_0 | 9.3 GB | 10.2 GB | — | ~9.4 | −0.1% ppl | 3.2 GB over |
| Q6_K | 7.2 GB | 8.0 GB | — | ~21 | −0.4% ppl | 1.0 GB over |
| Q5_K_M | 6.2 GB | 7.1 GB | 6K | ~46 | −0.8% ppl | 0.1 GB over |
| Q4_K_M | 5.3 GB | 6.1 GB | 35K | 59 | −1.9% ppl | Recommended |
| Q3_K_M | 4.3 GB | 5.1 GB | 67K | 72 | −5.4% ppl | Long context |
| Q2_K | 3.7 GB | 4.5 GB | 87K | 84 | −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 8 of its 32 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-1.5-9B:Q4_K_M \
-c 8192 -ngl 99
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.
01Download is 5.3 GB. Keep it on an SSD — a first load off a spinning disk takes minutes.
02Close anything else holding VRAM. A browser with hardware acceleration can sit on 1–2 GB.
03Only 0.9 GB is spare, so a long system prompt can still push it over. Its real ceiling here is 35K context.