The installer detects your GPU and first assigns a deterministic hardware tier. Linux and macOS then run the versioned catalog selector (ods/scripts/select-model.py), while Windows uses the PowerShell catalog selector in ods/installers/windows/lib/tier-map.ps1; both read ods/config/model-library.json to choose the best installable GGUF for the detected memory envelope. The final choice is written to .env as LLM_MODEL, GGUF_FILE, MAX_CONTEXT, and MODEL_RECOMMENDATION_*.

MODEL_PROFILE=qwen is the default catalog profile. MODEL_PROFILE=gemma4 forces Gemma 4 where available, and MODEL_PROFILE=auto uses Gemma 4 on NVIDIA, AMD Strix Halo, Apple Silicon and Intel Arc tiers. Override the model family with MODEL_PROFILE=gemma4 ./install.sh. You can switch models at any time after installing (see Switching Models).

The tables below are what the catalog selector picks today for common hardware with the default profile and a 64K context floor (python ods/scripts/select-model.py --profile qwen --installable-only --min-context 65536 …). Exact picks can differ with detected VRAM/RAM, host architecture and existing downloads. Measure throughput on your own machine after the first launch.

NVIDIA

| Memory | Default pick | Context | Example hardware | |--------|--------------|---------|------------------| | 8–16 GB VRAM | Qwen3.5 9B (Q4_K_M) | 64K | RTX 4060, RTX 3060 12GB, RTX 4070-class | | 24–32 GB VRAM | Qwen3.5 27B (Q4_K_M) | 64K | RTX 4090, RTX 5090 | | 48 GB VRAM | Qwen3.6 35B-A3B (UD-Q4_K_M) | 128K | RTX 6000 Ada, L40S | | 96 GB+ VRAM (x86_64) | Qwen3 Coder Next (Q4_K_M) | 128K | RTX PRO 6000, multi-GPU | | 128 GB unified (arm64) | Qwen3.6 35B-A3B (UD-Q4_K_M) | 128K | DGX Spark / GB10 |

AMD Strix Halo (Unified Memory)

| Memory | Default pick | Context | Hardware | |--------|--------------|---------|----------| | 64–128 GB unified | Qwen3.6 35B-A3B (UD-Q4_K_M) | 128K | Ryzen AI MAX+ 395 |

The selector routes unified-memory hosts away from Qwen3 Coder Next because of documented correctness issues on those backends.

Apple Silicon (Unified Memory, Metal)

| Memory | Default pick | Context | Example hardware | |--------|--------------|---------|------------------| | 8 GB | NVIDIA Nemotron3 Nano 4B (Q4_K_M) | 64K | M1/M2 base | | 16–36 GB | Qwen3.5 9B (Q4_K_M) | 64K | M4 Mac mini, M3 Pro | | 48 GB+ | Qwen3.6 35B-A3B (UD-Q4_K_M) | 128K | M4 Pro/Max, Mac Studio |

CPU only

| System RAM | Default pick | Context | |------------|--------------|---------| | 8–12 GB | Qwen3.5 2B (Q4_K_M) | 64K | | 16–20 GB | Qwen3.5 4B (Q4_K_M) | 64K | | 24 GB+ | Qwen3.5 9B (Q4_K_M) | 64K |

Intel Arc (Linux, SYCL)

Intel Arc is experimental (Tier C). Hardware detection does not identify Arc GPUs yet, so pass the tier yourself: ./install.sh --tier ARC_LITE for 6–8 GB cards or --tier ARC for 16 GB cards. See the Intel Arc guide.

---