Ajax AI model system requirements (estimates)
Ajax has not been released, so nothing on this page is a confirmed spec. Everything below is an estimate derived from the 9-billion-parameter Qwen 3.5 9B base model. Treat it as "what to expect", not "what to buy".
VRAM estimates by precision
| Build | VRAM (estimate) | Example hardware that fits |
|---|---|---|
| Full precision (bf16/fp16) | ~22 GB (reported analysis of the base model - sudoreview) | 24 GB cards: RTX 3090, RTX 4090, A10, L4 |
| 8-bit quantised | ~11 GB (reported analysis) | 12 GB cards, tightly: RTX 3060 12 GB, RTX 4070 |
| 4-bit quantised | ~5–6 GB (our estimate: standard 9B rule of thumb, ~0.6 GB per billion params + overhead) | 8 GB cards; 16 GB unified-memory Macs |
These are weights-only figures for the base model. Context (KV cache), runtime overhead and any always-on agent workload add memory on top - how much is unknown until release.
System RAM and CPU
If you run the model on GPU, system RAM requirements are modest - enough to load the file and run the host OS plus Odysseus. Offloading layers to CPU is technically possible with most local runtimes but will be slow for a 9B model; no official guidance exists yet. All of this paragraph is an estimate.
Don't forget: Odysseus itself has to run
Ajax is built to run inside Odysseus, the self-hosted workspace - the model handles the thinking, but Odysseus runs the agents, search, email and calendar around it. The Odysseus repo ships adocker/ folder for its services (GitHub), so plan for Docker and its own (unpublished for Ajax, but modest by AI standards) footprint on top of the model's VRAM. See how to run it locally for the guide skeleton.
What is still unknown
- Released formats: GGUF, AWQ, MLX, safetensors - none announced
- Context length of the fine-tune
- Which runtimes will be supported (Ollama, LM Studio, vLLM, llama.cpp)
- License, and whether it permits commercial or hosted use
- Whether smaller/larger variants will ship
When the files drop we will replace the estimates on this page with measured numbers. Until then, the practical answer is: any GPU with ≥8 GB will very likely run some quantised build of a 9B model - estimate, not a promise.
Get pinged the day Ajax ships
One email when the status on this page changes to released. Nothing else.
Stored only to send the release notification. See privacy.
Related: download status, FAQ, release status.