Supported models
What hf2q converts and serves.
See which model families hf2q converts and serves—and which capabilities each one exposes.
Model families
Each family lives in its own module inside hf2q's arch-registry — the single source of truth for tensor catalogs, quality thresholds, and smoke prompts.
| Family | Architecture | Serves | Notes |
|---|---|---|---|
| Gemma 4 | Dense + MoE | Chat completions with tools, streaming, and grammar-constrained sampling | Launcher: scripts/serve_gemma4_opencode.sh, default port 8082 |
| Qwen 3.5 / 3.6 / 3.8 | Dense + MoE + multi-token prediction | Chat completions with tools and streaming; image inputs with a bound vision projector (serve --mmproj) | Launchers: scripts/serve_qwen36_opencode.sh and serve_qwen38_opencode.sh (3.8-27B multimodal), default port 8081. Convert disk floor: 100 GB dense, 150 GB MoE |
| DeepSeek-V4-Flash-0731 | Compressed-attention MoE | Chat completions with tools and streaming | Launcher: scripts/serve_deepseek4_opencode.sh, default port 8081. Run one large family at a time. Embeddings and image inputs are refused explicitly |
| BERT / Nomic-BERT | Encoder, embedding-only | Embeddings via /v1/embeddings |
Quantization
hf2q convert accepts two families of --quant values. Anything else fails with a typed, actionable error.
Standard llama.cpp ftypes
All supported families
f32f16bf16q4_0q4_1q5_0q5_1q8_0q2_kq3_k_sq3_k_mq3_k_lq4_k_sq4_k_mq5_k_sq5_k_mq6_kiq4_nlByte-identical to stock llama-quantize output for the same ftype.
APEX algorithmic tiers
MoE architectures only
apex-qualityapex-i-qualityapex-balancedapex-i-balancedapex-compactapex-i-compactapex-miniI-tier variants require imatrix data via --imatrix or --imatrix-corpus.
Reserved names (dwq, unqualified apex, tq1_0 / tq2_0) fail with typed, actionable errors.
The support boundary
What the supported surface requires, and where it ends.
Platform
macOS on Apple Silicon (M1 or newer). The inference path is Metal-only, and the Rust toolchain is pinned at 1.88.0.
Disk floors
Conversion enforces per-architecture disk floors — 100 GB for Qwen 3.5 dense, 150 GB for Qwen 3.5 MoE. Preflight refuses to start below the floor plus 10 GB.
Fail-closed support
The published support contract is limited to families on this page, and reserved quant names fail with typed errors.
Serving surface
One OpenAI-compatible endpoint per launched family: /v1/chat/completions, /v1/embeddings, and /v1/models, with streaming SSE and a persistent block-prefix KV cache.