CapriChooms/CapriCore

These planning figures cover one conversation with 32K context and the bundled quantizations. VRAM figures describe model consumption, while system RAM targets describe total PC capacity. They do not include optional speech.

| Text model | Estimated model VRAM in use | System RAM target with a suitable GPU | Conservative CPU-fallback RAM target | | --------------------- | --------------------------- | ------------------------------------- | ------------------------------------ | | Qwen3.5 0.8B | 1.3–2.2 GiB | 8 GiB | 8 GiB | | Qwen3.5 2B | 2.8–4 GiB | 8 GiB | 16 GiB | | Qwen3.5 4B | 4–6 GiB | 12 GiB | 24 GiB | | Qwen3.5 9B | 7–10 GiB | 16 GiB | 32 GiB | | Ministral 3 3B (BETA) | 3.5–6 GiB | 12 GiB | 16 GiB | | Gemma 4 E2B (BETA) | 8–12 GiB | 16 GiB | 24 GiB |

GPU acceleration uses the installed llama.cpp Vulkan backend and a compatible driver. A card needs additional headroom for the display, other applications and any GPU voice. CPU execution and partial offload can be slower and consume more system RAM.

16 GiB RAM plus 16 GiB VRAM does not guarantee comfortable 9B operation. The 16 GiB RAM target assumes text-only use and a working GPU backend; it is not a promise that RAM stays almost unused. Windows, the browser, Capri's core, CPU speech recognition and optional TTS still need system RAM. Loading peaks, longer context and driver-managed shared memory can increase it further. On a 16 GiB PC, start with 2B/4B and speech off, or use an external API. Confirm actual offload rather than inferring it from VRAM capacity. Diagnostic reports include resources.json with the local model's resident CPU/GPU allocation counters and Capri's process working sets, separately from total system usage.