Open-weight models finally do real agent work — tool calls, multi-step plans, MCP. Arg0 points your agents at Ollama, LM Studio, or Apple's on-device models: same runtime, same workflows, zero per-token cost.
The landscape
A year ago "local model" meant a chatbot that hallucinated tool syntax. The ecosystem changed on four fronts:
Tool calling went standard
Open-weight models are now trained for function calling, not just chat — Qwen3's native template, Hermes' <tool_call> tags, gpt-oss's OpenAI-format calls. Benchmarks like BFCL, TAU2-Bench and MCP-Atlas now rank models on agentic tool use.
Laptop-class is good enough
Mixture-of-experts architectures (Qwen3-A3B, gpt-oss-20b at ~3.6B active params) plus 4-bit GGUF quantization put serious reasoning inside 8–16 GB of memory — unified-memory Macs included.
A real ecosystem of models
Qwen3, Nous Hermes 4.x, Llama, gpt-oss, GLM-4.7-Flash, Gemma, Ministral, Granite — most under Apache-2.0 or MIT licences, downloadable and redistributable.
Alignment you choose
Hermes-style models are trained for neutral alignment and steerability — your system prompt wins. Pick the model whose behaviour fits the job, not the vendor's default.
Worth pulling first
qwen3Safe default — native tool template + optional thinking, 0.6B to 235B
hermes3 / Hermes 4.xNous' steerable, neutrally-aligned agent line — hermes3 via Ollama, 4.x GGUFs via LM Studio or vLLM
gpt-oss-20bOpenAI's open weights — function calling + structured outputs
llama3.1:8bThe most-pulled tool-capable model — the reference implementation
GLM-4.7-FlashMIT-licensed MoE — strong τ²-Bench score at ~3B active params
In Arg0
Local models are first-class providers — not a workaround. Agents, tasks, and workflows don't know or care where the model runs.
Ollama
Detected automatically once installed. Arg0 reads the `tools` capability tag, so tool-capable models — qwen3, llama3.1, hermes3, gpt-oss — just work as agents.
LM Studio / llama.cpp / vLLM
Any OpenAI-compatible endpoint connects in Settings → AI Preferences — including GGUFs like Hermes 4.3 that you download straight from Hugging Face.
Apple Intelligence
On-device Foundation Models on macOS, exposed inside Arg0 as a native agent runtime with real tool calls — no install at all.
Hermes models work too. Nous ships its own Hermes Agent app, but the models themselves are open weights — pull hermes3 in Ollama or load a Hermes 4.x GGUF through LM Studio or vLLM, and it shows up in Arg0 as a first-class agent model with tool calls and MCP access like any other.
Where it shows up
Every place Arg0 takes a model, it takes a local one — the UI doesn't change, only where the inference happens.
Managed AI
Arg0 credits — works out of the box
Ollama
localhost:11434 — 3 models found
LM Studio
localhost:1234 — OpenAI-compatible
Apple Intelligence
on-device Foundation Models
Your API keys
OpenAI · Anthropic · Gemini · DeepSeek
AI Providers — local servers sit next to Managed AI and BYOK as equals.
PROVIDER
MODEL
MCP TOOLS
Per agent — provider, model, and MCP tools are picked independently for each agent.
Summarize this week's notes
agent: @qwen3 · mcp: filesystem
Publish the changelog draft
agent: @claude · cloud
Per task — a agent: line in .tasks.md routes one card to a local model and the next to the cloud.
Why local
Private by default
Prompts, files, and run logs stay on your machine. Nothing is sent to a provider — there is no provider.
No per-token cost
Local models don't consume Arg0 credits or a provider subscription. Run agents all day for free.
Works offline
No network, no problem. Agents keep working on a plane, in a train, or behind a firewall.
Same agents, same workflows
Local models plug into the same agent runtime, MCP tools, and .workflow.md automations as cloud models.
Install a runner
Grab Ollama or LM Studio — or use Apple Intelligence, which ships inside the desktop app.
Pick it in Settings
Arg0 detects local servers automatically. Choose the provider and model per agent.
Assign work
Chat, dispatch tasks from .tasks.md, or wire it into a workflow — exactly like a cloud model.
# pull an agent-capable open model
$ ollama pull qwen3
$ ollama pull hermes3 # Nous' steerable agent line
# both appear in Arg0's provider list
$ # agent: @qwen3 in a .tasks.md card
# the whole run stays on this machine
✓ 14 steps · 3 tool calls · 0 tokens billed
Managed AI, your own API keys, or a model on your desk — switch providers per agent, per task, per workflow.