Choose models

You don’t have to guess which model to run. hal0 ships a curated catalogue — a hand-picked list of good defaults, each pointing at a specific GGUF (or safetensors) file in a Hugging Face repo, sized for the target’s unified-memory pool. Pull any of them by id:

hal0 model pull qwen3.6-27b

The catalogue ships with each release (it lives in code, not a remote manifest), so you can never land on a stale pick list.

The Models view renders each curated entry as a card with size, license, and minimum-memory badges.

Chat picks

These are surfaced first by the FirstRun wizard, ordered roughly largest to smallest. Most of the headline picks use Unsloth’s UD-Q4_K_XL dynamic quant or plain Q4_K_M for the size/quality sweet spot; qwen3.6-27b ships a ROCmFP4 pre-quantized build tuned specifically for Strix Halo instead.

Id	Model	Size	Min memory	License	Notes
qwen3-coder-next	Qwen3 Coder Next	~49 GB	~56 GB	Apache-2.0	Frontier coding model; needs the full pool.
qwen3.6-27b	Qwen3.6 27B (MTP ROCmFP4)	~17 GB	~22 GB	other (see HF repo)	General-purpose default with MTP and vision support; ROCmFP4 build for Strix Halo.
gpt-oss-20b	GPT-OSS 20B	~12 GB	~16 GB	Apache-2.0	OpenAI open-weights 20B.
qwen3.5-9b	Qwen3.5 9B	~6 GB	~8 GB	Apache-2.0	Lean default; fits alongside embed + voice slots.
qwen3.5-0.8b	Qwen3.5 0.8B	~0.6 GB	~1 GB	Apache-2.0	Tiny; sub-second cold start. Doubles as the install smoke test.
qwen3-4b	Qwen3 4B Instruct	~2.5 GB	~4 GB	Apache-2.0	Fast all-rounder for a 4–8 GB budget.
llama32-3b	Llama 3.2 3B Instruct	~2 GB	~3 GB	Llama-3.2-Community	Small and fast; good for low-memory hosts.
phi3-mini	Phi-3 Mini 4K Instruct	~2.4 GB	~3 GB	MIT	Compact reasoning, MIT-licensed.

Tip

qwen3.5-0.8b is the recommended first pull on a fresh install — it verifies the whole slot lifecycle (pull → load → serve) in seconds before you commit to downloading a 20+ GB pick.

Community MTP builds behind the seed Stacks

The catalogue also carries exact HF coordinates for about a dozen community MTP builds that don’t show up as FirstRun wizard cards. Five of them are the agent/utility model pairs baked into hal0’s three built-in Stacks — saber, forge, and pi — so those Stacks resolve their models and pull them by id instead of failing on a coordinate-less registry row; the rest come from the same Strix-Halo bench roster those Stacks were tuned against. They’re still ordinary catalogue entries — hal0 model pull <id> works on any of them — just not curated for the first-run experience.

Id	Model	Size
qwen3-6-35b-a3b-nsc-ace-saber-mtp-f16-to-rocmfp4-strix-lean	Qwen3.6 35B-A3B NSC ACE SABER (ROCmFP4 STRIX)	~19 GB
chadrock3-6-27b-pi-agent-mtp-rocmfp4-strix-lean	CHADROCK3.6 27B Pi-Agent (ROCmFP4 STRIX)	~14.8 GB
chadrock3-6-35b-uncensored-mtp-strix-lean	CHADROCK3.6 35B Uncensored (MTP STRIX)	~19 GB
qwen3-6-35b-a3b-halostrix-dyn-mtp-v7	Qwen3.6 35B-A3B HaloStrix Dyn (MTP v7)	~22.6 GB
qwopus3-6-27b-v2-mtp-bf16-to-rocmfp4-strix-lean	Qwopus3.6 27B v2 (ROCmFP4 STRIX)	~14.8 GB
qwopus3-6-27b-coder-mtp-q6-k	Qwopus3.6 27B Coder (MTP Q6_K)	~22.4 GB
qwopus3-5-9b-coder-mtp-q6-k	Qwopus3.5 9B Coder (MTP Q6_K)	~7.6 GB
qwopus3-5-4b-coder-mtp-q6-k	Qwopus3.5 4B Coder (MTP Q6_K)	~3.6 GB
qwen3-5-9b-deepseek-v4-flash-mtp-q6-k	Qwen3.5 9B DeepSeek V4 Flash (MTP Q6_K)	~7.6 GB
gemma-4-12b-it-ud-q4-k-xl	Gemma 4 12B IT (UD Q4_K_XL)	~7.4 GB

All ten carry the mtp tag. The Gemma pick is also the reason hal0 pins any Gemma model on any q8-KV profile back to f16 KV automatically — see FAMILY_DEFAULTS for why.

Embed and rerank picks

llama.cpp-compatible GGUFs that fan out to gpu-vulkan / gpu-rocm / gpu-cuda / cpu. Assign these to an embedding or reranking slot.

Id	Model	Capability	Size	License
nomic-embed-text-v1.5-q8_0	Nomic Embed Text v1.5 (Q8_0)	embed	~0.15 GB	Apache-2.0
bge-base-en-v1.5-q4_k_m	BGE Base EN v1.5 (Q4_K_M)	embed	~0.07 GB	MIT
bge-reranker-base-q4_k_m	BGE Reranker Base (Q4_K_M)	rerank	~0.26 GB	MIT
bge-reranker-v2-m3-q4_k_m	BGE Reranker v2 M3 (Q4_K_M)	rerank	~0.44 GB	Apache-2.0

Note

Embedding quality is brittle under aggressive quantization, so the nomic default ships at Q8_0 — the size delta is negligible on a unified pool. Rerank slots must be created with --type reranking (see Manage slots).

Image picks

Routed through ComfyUI — these land in ComfyUI’s models/checkpoints tree rather than the per-id model directory, so its loaders pick them up directly.

Id	Model	Size	License
sdxl-turbo	SDXL Turbo	~6.5 GB	SAI-NC-Research-Community (research only)
sdxl-lightning	SDXL Lightning (8-step)	~6.8 GB	CreativeML-OpenRAIL++
sd-1.5-pruned-emaonly	Stable Diffusion 1.5	~4.3 GB	CreativeML-Open-RAIL-M

Caution

Image picks span the licensing spectrum on purpose — SDXL Turbo is research-only, SD 1.5 is RAIL-M. Read the license before using output commercially.

The dashboard’s Models view keeps these off the dispatcher-facing Models tab entirely — they live on their own Image / ComfyUI tab, grouped by the ComfyUI models-tree subdirectory each one actually lands in (checkpoints, loras, vae, upscale_models, …). See Generate images for how that surface works.

How model fit is evaluated

Before a model is assigned to a slot, hal0 computes a fit verdict — allowed, degraded, or blocked — with stable reason strings the dashboard renders as chips. The check runs three gates in order:

1. Type match. The model’s classified modality must match the slot type. A chat model maps to an llm slot, embed → embedding, rerank → reranking, stt → transcription, tts → tts, image → image. A mismatch is blocked (model.slot_type_mismatch).

2. Resolvability. The model id must resolve — present in the local registry. An unresolvable id is blocked (model.not_resolvable).

3. Profile + device class. The slot’s profile must support the slot type — a mismatch (e.g. a tts-only profile on an llm slot) is blocked (profile.unsupported_slot_type) — and the device must agree with the profile’s device class (gpu-rocm/gpu-vulkan → gpu, cpu → cpu, npu → npu). NPU and image mismatches are hard blocked failures (profile.device_class_mismatch) — they route to different runtime families. A GPU↔CPU mismatch reports the same reason string but as degraded instead: a custom image may still run, but it needs operator attention before launch.

A degraded verdict is allowed to proceed; only blocked stops the assignment. This is why the device you pick for a slot has to be consistent with its profile — see Manage slots for the --hardware flag that sets it.

Tip

Fit-checking runs at assignment time. If a slot’s already-configured default later stops resolving locally — its file was removed, or a seed pinned an id that was never actually pulled under that exact name — hal0 slot load doesn’t crash-loop on the missing path. load() falls back to the best locally-registered model that matches the slot’s capability, preferring whichever candidate shares the most leading name tokens with the configured id (then the largest one on disk), and always excluding diffusion/image/video artifacts so a text slot can never load one by accident. The substitution is logged (slot.model_default_fallback) so you can go fix the slot’s configured default.

Off-catalogue models

Anything not in the catalogue still works — search Hugging Face, inspect a repo’s variants, and pull by coordinates. See Pull and register models.