Open-source home AI inference platform

Stop running models from a chat tab.

hal0.dev · Install · Docs · Benchmarks · Roadmap · Discord

---

hal0 turns a Linux box — ideally a Ryzen AI Max+ 395 with 128 GB of unified memory — into a private, OpenAI-compatible inference appliance. Chat, completions, embeddings, reranking, transcription, speech and image generation all answer on one :8080/v1 surface, on hardware you own, with no per-token bill.

Every inference workload runs as its own podman container under a hal0-slot@<name>.service unit. hal0-api on :8080 is the sole control plane — it owns slot state machines, dispatches /v1/* requests to the right slot port, and serves the dashboard. No shared inference daemon; no extra process to babysit.

curl -fsSL https://hal0.dev/install.sh | sudo bash

⚡ The installer is the whole setup. There is no separate first-run wizard afterwards. install.sh asks where models should live, optionally takes a HuggingFace token, seeds every capability slot, downloads the hal0 brain steward model, and starts the API. When it finishes, open the dashboard and assign models to slots.

Status: release candidate — v1.0.0-rc.5. v1.0.0 stable is the target of the current cut. rc.5 is the validation candidate: it closes every defect the rc.4 fresh-install sweep found on real hardware. See CHANGELOG.md.

Screenshots

Dashboard — unified-memory hero, at-a-glance health, and live slot rows.

Slots — the inference engine: per-slot models, devices, and live telemetry.

Model registry — catalogue with quant chips and pull status.	Image generation — ComfyUI queue and GPU gauges.	Memory — Hindsight facts and graph extraction.
Services — Open WebUI, ComfyUI, Hermes, Hindsight, n8n with mDNS.	Stacks — declarative slot + profile + model bundles, applied atomically.	Settings — full hal0.toml parity with an Advanced section.

_{More in the docs.}

Why hal0

Point your editor at your own box. Aim any OpenAI-compatible client at http://<box>:8080/v1 — Claude Code, Cline, Continue, an SDK, a cron job. A dedicated coder slot can serve your editor while the agent slot serves chat, each on its own container, model and flag set.

Retrieval grounded on your data. Embeddings and reranking are first-class slots, not an afterthought: /v1/embeddings and /v1/rerank run on the same box as chat, and the bundled Hindsight memory engine turns conversations into recallable facts behind an MCP server.

Speech and images on the same control plane. Transcription (/v1/audio/transcriptions), speech (/v1/audio/speech) and image generation (/v1/images/generations, /v1/images/edits) are the same API, the same auth posture, the same dashboard — no second stack to operate.

Not another llama-server wrapper. Every workload is a real systemd-managed slot with a typed lifecycle, persisted state (/var/lib/hal0/slots/<name>/state.json), SSE status streaming, per-slot journald logs and a GPU arbiter that keeps image generation from stealing the iGPU out from under a streaming completion.

Strix Halo native, but not Strix-Halo-only. The probe is UMA-aware on Strix Halo and falls back to portable parsers (/proc/cpuinfo, /proc/meminfo, lspci) everywhere else. platform on /api/hardware resolves to one of strix-halo, wsl2, lxc, proxmox-kvm, kvm, bare-metal-{nvidia,amd,intel}-gpu or bare-metal-cpu-only, and the dashboard only labels memory "unified" when it actually is. Slot-fit warnings size against the real unified pool, not a BAR carve-out.

NPU multi-role via the FLM trio. On Strix Halo with FastFlowLM installed, a single FLM process inside the hal0-toolbox-flm container hosts chat + transcription + embedding concurrently on one AMDXDNA hardware context — ~2 GB NPU memory, gemma3:1b at ~38.6 tok/s + Whisper-V3-Turbo + Embedding-Gemma all coresident. hal0 exposes this as one seeded flm slot (chat-first) whose [npu] table — or the dashboard's NPU drawer — toggles ASR and embed on the running flm serve. The host-side FLM .deb is installed for device-sanity probes only; inference runs in the container. See docs/concepts/strix-halo.mdx.

An agent that lives on the box. The hal0-brain steward drives the platform from the dashboard's top bar, and Hermes/Pi install as real services prewired to the local API and MCP servers.

Reliability bar. Atomic env-file writes. Schema-validated TOML. Structured error envelopes ({"error":{"code":"slot.not_ready",...}}). Cosign-verified self-update with one-flag rollback. Per-type LRU concurrency with active-inference protection — a serving slot cannot be evicted out from under a streaming request. Privileged operations run through an allow-listed root seam, never a free-form shell.

Benchmarks

Numbers below are from the public leaderboard at hal0.dev/benchmarks — 26 models measured end to end on the reference box (AMD Ryzen AI Max+ 395, Radeon 8060S iGPU, 128 GB LPDDR5X, ROCm 7.2.4), rocm lane, decode and prefill throughput plus TTFT, speculative accept rate, memory, watts and temperature per run.

Model	Params	Quant	Decode (tok/s)	Prefill (tok/s)
qwen3.5-0.8b	0.8B	q4	169.8	6,248
chadrock3-6-35b-uncensored	35B MoE	q4	102.1	890
chadrock-35b-ace-saber	35B MoE	f16	100.5	903
qwen3.6-35b-a3b-crown	35B MoE	f16	84.4	873
qwopus3-5-4b-coder	4B	q4	85.0	889
qwen3-4b	4B	q4	61.9	1,849
gemma-4-26b-a4b	26B MoE	q4	40.9	1,336

Concurrency and non-LLM lanes, same box:

Workload	Measurement
Chat + embed, concurrent	~258 tok/s aggregate
NPU chat (FLM trio, gemma3:1b)	~38.6 tok/s in ~2 GB NPU memory
TTS — Kokoro (CPU)	~0.18 real-time factor
TTS — Qwen3-TTS (ROCm)	~0.48 real-time factor (~2.1× realtime)

Benchmarking is a first-class subsystem, not a spreadsheet: hal0 bench plan|run|results|eval|bundle|upload drives llama-bench/llama-server under a record schema with a dedup identity key, an in-dashboard Benchmarks page (roster / runs / evals / run-queue), and regression journaling. Results never leave the box on their own — hal0 bench bundle writes a content-addressed, host-redacted archive you can attach anywhere, and hal0 bench upload publishes one only when you type it. See docs/reference/model-roster-benchmark.mdx.

What's in the box

• OpenAI-compatible /v1/* API — chat, completions, embeddings, rerank, audio transcriptions, audio speech, image generations, image edits, models. Drop-in for any OpenAI SDK; point your client at http://localhost:8080/v1 and go.
• Slots — each named target carries a type (llm | embedding | reranking | transcription | tts | image), a device (gpu-rocm | gpu-vulkan | gpu-cuda | cpu | npu | img), a model, plus enabled and an optional default. Ten curated slots (agent, brain, coder, embed, flm, img, qwen3tts, rerank, tts, utility) are seeded into /etc/hal0/slots/<name>.toml on install and each gates on its own runtime validation at load time — the flm NPU slot simply stays grey without FastFlowLM hardware. Add your own with hal0 slot create NAME --type TYPE --model MODEL. A deliberate hal0 slot delete is tombstoned, so re-running the installer never resurrects it.
• Model owns its tune — since v1.0 the model, not the slot, owns context_size, extra_args, chat_template, profile and MTP. A slot's [model].context_size is a ceiling; a model with no stamped tune inherits its profile's template as a floor.
• Container runtime — every slot runs as its own podman container under hal0-slot@<name>.service, bound to a loopback port (8081–8099 + fixed seeds). hal0-api on :8080 is the only control plane. See docs/concepts/architecture.mdx.
• Provisioning lives in the installer — install.sh is the single user-facing entry point: model storage, optional HuggingFace token (on a terminal only), curated slot seeds, the brain steward pull, an optional agent-model offer, and service start. HAL0_SKIP_SETUP=1 skips first-run seeding, HAL0_SKIP_BRAIN_MODEL=1 skips the brain pull, and HAL0_NONINTERACTIVE=1 suppresses both prompts. See installer/README.md for the full step list and env-var table.
• Hardware-aware probe — detects GPU / NPU / unified memory, writes /etc/hal0/hardware.json, and surfaces VRAM/RAM fit warnings inline in the slot form and during install.
• Dispatcher — registry-aware routing, single-flight, cold-cache prefetch, and upstream fallback (OpenRouter, Anthropic, OpenAI, Google AI Studio, Ollama, custom OpenAI-shaped endpoints). Mix local + remote per-model in one config.
• Dashboard — React 18 + Vite UI for slot/model management, hardware-aware configuration, live logs, system health and a built-in chat page (popout window + reasoning toggle). SSE-backed status and log tail; a Journal panel streams per-slot container logs from journald. Live telemetry — Power & Thermal (GPU clock/temp/power) and per-slot throughput — is on by default. The slots page splits into Inference | Image Gen tabs; the Image-Gen tab operates the ComfyUI container (live GTT/RAM gauges, queue depth, model inventory) with a gated inference ⇄ generation iGPU switchover behind a blast-radius confirm.
• AI Capabilities page — one settings surface for TTS, STT, embeddings, reranking, image generation and the NPU, with per-panel probe-failure surfacing instead of a silent grey Save.
• Stacks — declarative, runtime-switchable model/slot layouts. A StackConfig is planned into a change set, applied atomically (with rollback) and converged against the live slot set; drift is detected by content hash. Stacks export/import via a checksummed .hal0stack.json envelope, and a live config can be snapshotted back into a Stack.
• OmniRouter (8 tools) — generate_image, edit_image, text_to_speech, transcribe_audio, analyze_image (vision), embed_text, rerank_documents, route_to_chat. Dispatched client-side from any chat slot whose model carries the tool-calling label, and filtered per request — a tool only appears in the prompt if its target slot is enabled and its model carries the required labels.
• Companion-service management — /api/services is the declarative source of truth for OpenWebUI, Hermes, Hindsight, ComfyUI and n8n: audit-logged systemd lifecycle actions (start/stop/restart/enable/ disable), mDNS .local advertisement, and a dedicated Services page. OpenWebUI is prewired at :3001 with zero config.
• Image generation, day one — POST /v1/images/generations via ComfyUI in the img slot container (ROCm). The installer seeds the img slot and its ComfyUI model share.
• Atomic self-update with rollback — hal0 update --channel stable|preview|nightly. Cosign-verified tarballs swap a /usr/lib/hal0/current symlink; --rollback reverts. Staging happens in a root-only directory with the authenticated digest pinned to the very file object the archive is extracted from.
• One-line install — curl -fsSL https://hal0.dev/install.sh | sudo bash. The only model it downloads is the brain steward; every other slot lands model-less for you to fill. Piped through bash it never prompts. (--models-dir=PATH or HAL0_MODELS_DIR=PATH redirects model pulls off /var/lib/hal0/models.) The bootstrap requires jq and cosign, authenticates the exact channel manifest before strict schema/channel parsing, then sha256- and Sigstore-bundle-verifies the tarball before handing off to installer/install.sh. Re-running it is safe: every provisioning step is idempotent and existing slot configs are never overwritten.
• One-line Proxmox VE install — on a Proxmox host, bash -c "$(curl -fsSL https://raw.githubusercontent.com/Hal0ai/hal0/main/scripts/proxmox-ve/hal0.sh)" creates an unprivileged Debian 13 LXC and runs the standard bootstrap inside it. --advanced opens whiptail prompts; every parameter has an env-var override (CTID, RAM_MB, STORAGE, …). Hardware-agnostic — Strix Halo passthrough still needs the privileged-LXC recipe. See scripts/proxmox-ve/README.md.

Agents, MCP, and the brain steward

hal0 ships two MCP servers and two bundled agents. The MCP servers (/mcp/admin for slot / model / capability / config / hardware / log admin, /mcp/memory for Hindsight-backed long-term memory) are reachable by any MCP-speaking client — Claude Code, RAG services, your own scripts. Both bundled agents can be installed simultaneously: pi-coder (CLI shape, from the Hal0ai/pi-mono fork via @earendil-works/pi-coding-agent on npm) and Hermes-Agent (service shape, via the hal0-owned hermes wrapper — hal0-hermes remains a back-compat symlink — connecting to hal0-api at HAL0_INFERENCE_BASE=http://127.0.0.1:8080). Install either with hal0 agent install <name>. Destructive MCP calls (model_pull, slot_delete, config_write, …) gate through a header bell + inbox modal in the dashboard, with CLI parity via hal0 agent approvals {list,approve,deny}.

The hal0-brain steward is a third, built-in persona wired to the dashboard's top-bar agent chat — no install step. It drives the platform through the 180-tool hal0-admin catalog (models, slots, stacks, profiles, settings, upstreams, benchmarks) under a per-persona tool policy: tools_allowed hides tools, require_approval tightens them, and a POLICY_NO_LOOSEN floor keeps destructive and secret-bearing tools (*_delete, config_write, provider_credential_write) always gated — no persona edit can disarm them. Gated calls pause the turn for inline approve/deny. A [brain_chat] config gives operators a hard kill switch, a read-only mode, loop-budget knobs and a slot override (run the steward on any slot, e.g. hal0/npu) from Settings → Agents / Brain. The shipped brain models are native tool-callers, so the steward executes its own calls on the brain slot out of the box; binding a larger model to the agent slot is the optional upgrade path via the [brain_chat] tool_model fallback.

See docs/concepts/agents.mdx, docs/reference/mcp-tools.mdx and docs/guides/run-agents.mdx.

Backends

Each capability runs in its own container, supervised by hal0-slot@<name>.service. The profile in /etc/hal0/profiles.toml pins the flag bundle; the image is slot-owned.

Capability	Profile / image	Device	Notes
chat + embed + rerank	rocm / rocm-dnse / rocm-moe / vulkan / cuda (experimental)	ROCm / Vulkan / CUDA / CPU	ROCm FP4 fork baked into the image; MTP via --spec-type draft-mtp on the rocm-dnse/rocm-moe profiles
chat + STT + embed (NPU)	flm (hal0-toolbox-flm)	AMD XDNA (opt-in)	FLM trio: one container, [npu] asr/embed toggles, ~2 GB NPU mem
transcription	stt (hal0-toolbox-moonshine) / FLM trio above	CPU / AMD XDNA	voice.stt switches Moonshine (CPU) ⇄ FLM's whisper-v3:turbo (NPU) without reconfiguring the slot; no GPU STT engine
TTS	tts (hal0-toolbox-kokoro) / tts-qwen3 (hal0-toolbox-qwen3tts)	CPU / ROCm	voice.tts switches Kokoro (CPU) ⇄ Qwen3-TTS (GPU) without reconfiguring the slot
image	comfyui (docker.io/kyuz0/amd-strix-halo-comfyui)	ROCm	Exclusive GPU via arbiter; SD Turbo / Flux-2-Klein-9B

Every seeded profile requests -fa auto, so Flash Attention is probed at startup and falls back cleanly when it cannot be scheduled on its layer's device. A model with a measured win from forcing it can set -fa on in its own defaults.extra_args.

The NPU path is opt-in: the installer places a FastFlowLM .deb on the host (SHA-pinned per distro — Ubuntu 24.04/25.10/26.04 and Debian 13) for device-sanity probes (flm validate); inference runs inside the hal0-toolbox-flm container. With FLM present the installer seeds the flm slot, and the NPU trio activates once flm validate passes.

For the container-runtime operator reference — service layout, slot TOML fields, profiles, GPU arbiter and day-2 commands — see docs/concepts/architecture.mdx and docs/guides/manage-slots.mdx.

Hardware

Linux + systemd is the only hard requirement (installer/lib/preflight.sh, preflight_systemd). macOS and Windows are not in scope for v1.

Tier	Hardware	Status
First-class	AMD Ryzen AI Max+ 395 ("Strix Halo") with iGPU + XDNA NPU + 128 GB unified	Reference deployment. All published perf numbers come from this box.
First-class	AMD Ryzen AI Max 385 / 390 with 64 GB unified	Same path; small + mid tiers fit, 70B Q4 with shorter context.
Experimental	NVIDIA RTX 30/40/50 (10–32 GB)	Dedicated cuda seed profile — upstream ghcr.io/ggml-org/llama.cpp:server-cuda via CDI (nvidia-container-toolkit), with multi-GPU gpu_index pinning on the gpu-cuda device. Auto-falls back to the vulkan profile when CDI isn't present. /api/backends doesn't yet auto-advertise gpu-cuda.
Supported	AMD Radeon RX 7000 / discrete (16–24 GB)	ROCm or Vulkan container profiles; same hal0-slot@<name> lifecycle.
Fallback	CPU-only x86_64	cpu-llm profile — the Vulkan toolbox image run CPU-only (no GPU passed to the container). Usable for tiny models / smoke tests, not the headline experience.

Setup and day-2 operation

After the one-liner finishes, the dashboard is at http://<box>:8080 and the API at http://<box>:8080/v1. Assign models from the dashboard, or:

hal0 model pull <hf-repo>/<file.gguf>   # resumable, disk-preflighted
hal0 slot load agent --model <id>       # bind a model and start the container
hal0 status                             # system + slot summary
hal0 chat --brain                       # terminal REPL against the steward

hal0 model pull streams from Hugging Face into registry.toml under the user.* namespace — a disk-space preflight fails fast before a multi-GB pull, and an interrupted pull resumes via HTTP Range instead of restarting from zero.

Diagnostics:

hal0 doctor all              # full pre-flight against the live host
hal0 doctor ports            # port claims + netavark DNAT rule audit
hal0 doctor ports --fix      # prune stale DNAT rules that black-hole a port
hal0 system-info             # host / GPU / NPU / runtime evidence (read-only)

Services and logs:

systemctl status hal0-api
systemctl list-units 'hal0-slot@*'
journalctl -fu hal0-api
journalctl -fu 'hal0-slot@*'
systemctl restart hal0-slot@agent    # restart a wedged slot container

hal0 uninstall [--keep-data] tears down a running install (a thin wrapper over installer/uninstall.sh).

Auth posture

hal0 ships built-in login, sessions and a deny-by-default route classifier, with three tiers — anon → client → admin. There is no per-user account system: a browser session is always admin-equivalent, and programmatic callers use Authorization: Bearer <key> (or ?api_key= for WebSocket upgrades) against an admin or client key.

Auth ships off by default. A fresh v1.0 install is trusted-LAN-open — the right posture for a homelab appliance — until an operator turns enforcement on:

hal0 auth rotate admin     # mint an admin key into /etc/hal0/api.env
hal0 auth require on       # enforce; takes effect on the next request
hal0 auth status           # show the current posture

hal0 does not terminate TLS or manage a user directory. If it is reachable from anything you don't physically control, front it with a reverse proxy that owns TLS — Traefik, nginx, Cloudflare Tunnel. Note that the optional OpenWebUI companion is a second listener on :3001 that this auth model does not cover. See docs/operate/auth.mdx and docs/concepts/security.mdx.

Proxmox integration (optional)

If hal0 runs inside a Proxmox LXC, the container only sees its own cgroup slice of memory — other tenants, ZFS ARC and the host kernel draw from the same physical DIMMs as GPU GTT but are invisible from inside. To surface that, drop a read-only PVEAuditor API token into the dashboard's Settings → "Proxmox integration" panel. The unified-memory bar then swaps to the physical host's DIMM total and adds a muted "Proxmox host" segment for other-tenant + kernel pressure. The token is sensitive, stored 0600 at /etc/hal0/proxmox.json, and never echoed back by the API. Bare-metal and VM installs leave the panel off.

Project layout

hal0/
├── src/hal0/         # Python package (FastAPI API + capability layer + ContainerProvider + CLI)
│   ├── providers/    # ContainerProvider, FLMProvider, ComfyUI, etc.
│   ├── slots/        # slot manager, state machine, GpuArbiter
│   ├── bench/        # benchmark planner, runner, store, publish
│   └── omni_router/  # client-side tool-calling loop + tool definitions
├── ui/               # React 18 + TypeScript + Vite + Tailwind 4 dashboard
├── installer/        # install.sh (writes /etc/hal0/, systemd units, hal0-api.service)
│   ├── etc-hal0/     # curated slot seeds + profiles.toml
│   └── systemd/      # hal0-agent@ template units
├── tests/            # pytest suite (α unit, β integration, γ release-gate)
│   └── release-validation/   # versioned RC validation kit (lanes, regressions, boxes)
├── docs/             # user docs (mirrored to hal0.dev/docs) + docs/adr/

The model catalog lives at /var/lib/hal0/registry/registry.toml — the single source of truth for HuggingFace coordinates, SHA-256 digests and curated filenames. Per-leaf symlinks under the models directory redirect to a separate storage volume when one is in use.

Quick start (development)

# backend
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
hal0 serve --reload

# frontend
cd ui
npm install
npm run dev

The API lives at http://127.0.0.1:8080, the dashboard at the Vite dev server URL (usually http://127.0.0.1:5173).

Roadmap

No dates — items are direction. The closer to the left, the closer to running on your box. Full version at hal0.dev/#roadmap.

Shipped (v1.0)

• Per-slot podman containers — every inference workload in its own hal0-slot@<name>.service; ContainerProvider + profiles.toml replaced the single-daemon model
• hal0-brain steward — top-bar agent chat driving the 180-tool hal0-admin catalog under a per-persona tool policy, pausing turns on gated tools for inline approve/deny
• Model-owned tuning — the model, not the slot, owns context size, MTP and launch flags; profiles supply a floor, never an override
• Stacks — declarative model/slot layouts applied atomically with rollback, drift detection and a portable export envelope
• In-dashboard benchmarks — the hal0.bench engine, /api/benchmarks, a Benchmarks page (roster / runs / evals / run-queue), shareable bundles and leaderboard publishing
• Upstream model controls — external providers (OpenRouter, Anthropic, OpenAI, Google AI Studio, Ollama, custom) with reactive CRUD, per-upstream model filters and an enabled kill switch across CLI, MCP and dashboard
• Companion services, one surface — Open WebUI, ComfyUI, Hermes, Hindsight and n8n from a dashboard Services page with mDNS discovery
• FLM trio NPU packing — chat + ASR + embed coresident on one AMDXDNA hardware context, toggled from the slot TOML's [npu] table
• GpuArbiter exclusive image mode — ComfyUI gets the iGPU to itself; LLM GPU slots pause and resume automatically
• OmniRouter client-side tool-calling — 8 tools, dynamic per-request filtering, route_to_chat cross-slot delegation
• GPU generalization — experimental CUDA and per-slot gpu_index pinning for multi-GPU hosts alongside the ROCm/Vulkan defaults
• Built-in auth — three tiers, deny-by-default route classification, key rotation and enforcement toggling from CLI or dashboard
• Privileged-seam hardening — the root wrapper allow-lists the content of every unit, drop-in and quadlet it writes; update staging is root-only with the verified digest pinned to the extraction handle
• Installer-owned first-run provisioning — install.sh is the single entry point: curated slot seeds, brain steward pull, service start
• registry.toml as sole model catalog, slot-state Prometheus + EventBus journal, hal0 mcp CLI with two-way MCP, cosign-keyless self-update with rollback, bundled agents (pi-coder / Hermes-Agent)

Soon

• Federated memory — recall across local + remote sources behind one memory surface (the MCP-client side shipped in v0.9)
• Loadout presets — curated model/slot presets you can flash onto a fresh install
• AUR PKGBUILD & Ubuntu PPA — native distro packages on top of the install script; pacman and apt as first-class install paths
• Light mode toggle

Exploring (v1.x +)

• Multi-host federation — a slot mesh across LAN boxes — primary on the Strix Halo, embed on the workstation, all behind one /v1/* surface
• Fine-tune & LoRA hot-swap — attach and rotate LoRAs against a warm base model without unloading the underlying weights
• Per-model rate limits & budgets — cost-style accounting for local inference; cap a chatty agent without taking the whole box down
• ChatOps adapters — Slack and Matrix bridges as extensions

License

Apache 2.0. See LICENSE.

Contributing

The contribution model is still being decided. File issues for discussion; PRs aren't being accepted from outside contributors yet. See CONTRIBUTING.md for the test tiers and the eventual flow. For questions and general chat, join the hal0 Discord.

This project adheres to the Contributor Covenant code of conduct. To report a security vulnerability, see SECURITY.md.