Skip to content

← All guides

ZeroClaw with Ollama and Qwen3.5: Tested Local Setup

ZeroClaw.net · Updated

Tested 16 October 2026: ZeroClaw v0.8.5, Ollama 0.35.0, qwen3.5:0.8b on an Apple M3 Mac with 8 GiB unified memory. We verified the endpoint, ran a CLI turn and a dashboard conversation, and recovered from a missing model. This is a small text-response test, not a coding, vision or tool-use benchmark.

ZeroClaw runs the agent; Ollama serves the local model. Their memory needs are separate. This revision replaces the old Q13-30B, provider = "llama", dummy-key and top-level config examples. It also removes the unsupported fork and two-GPU claims.

All images below are raw screenshots from this test, with no arrows, added text or generated replacements. Click the full-size links to read small UI fields. Commands remain copyable because terminal screenshots were unavailable; private pairing and credential screens are excluded.

1. Choose the exact model and install the runtimes#

Use the ZeroClaw installation guide and official Ollama download. Check zeroclaw --version and ollama --version. Our test used the official macOS arm64 ZeroClaw archive and retained its sibling web/dist dashboard directory. Rust compilation is not required for this prebuilt route.

Start with qwen3.5:0.8b for this reproducible smoke test. Qwen3 and Qwen3.5 are different families; do not substitute an invented name or an unspecified latest tag. The official catalog offers multiple sizes, but we did not execute the larger variants.

Official Qwen3.5 0.8B model page showing the exact run command and model details

Open full-size model screenshot. Source: Ollama's model tag.

2. Start Ollama and finish the download#

If Ollama is not already serving requests, run this in one terminal and keep it open:

ollama serve

In a second terminal:

ollama pull qwen3.5:0.8b
ollama list
curl -fsS http://127.0.0.1:11434/api/tags

Wait for the pull to finish. ollama list must contain qwen3.5:0.8b. Our downloaded GGUF variant reported Q8_0 quantization, 1,322,069,277 bytes on disk and this digest:

de63045f29755d4721c09d54c9b45c2dadb644f4fd7d604d4cac90d96049971c

Tags can change; record your own digest when reproducing a result. The catalog also has other backend variants, so use the installed model's metadata rather than assuming every download is identical.

Installed model inventory from the local Ollama API

Open full-size inventory screenshot. This is the actual local /api/tags response, formatted by the browser.

3. Configure the provider and agent#

Create a fresh directory named qwen-test and save the tested config as qwen-test/config.toml. It uses schema version 3:

schema_version = 3

[providers.models.ollama.local]
model = "qwen3.5:0.8b"
uri = "http://127.0.0.1:11434"
temperature = 0.0
num_ctx = 4096
num_predict = 128
native_tools = false

[agents.assistant]
model_provider = "ollama.local"
risk_profile = "reader"

[agents.assistant.workspace]
unrestricted_filesystem = false

[risk_profiles.reader]
level = "readonly"
workspace_only = true
allowed_tools = ["time"]

ollama.local is the provider alias referenced by the agent. No dummy API key is needed for this local Ollama server. native_tools = false selects the non-native tool path; the prompt below asks for no tools. This test does not demonstrate tool reliability or force thinking off. The old [provider_extra_body] example is not part of this tested config.

zeroclaw --config-dir ./qwen-test config migrate --json

Expected result: "valid": true, "schema_version": 3, "error": null. Read the configuration guide before migrating an existing installation.

In the dashboard, Config → Model providers → ollama / local → Connection shows the URI. The screenshot uses the field filter to make the endpoint easy to find:

ZeroClaw provider connection settings showing http://127.0.0.1:11434

Open full-size endpoint screenshot. The Model tab contains the model identifier:

ZeroClaw model field showing qwen3.5:0.8b

Open full-size model settings screenshot.

For native ZeroClaw and Ollama on the same machine, use the loopback endpoint above. Inside Docker, loopback refers to the container; follow the Docker guide to choose a reachable host endpoint. Do not expose Ollama to the public internet just to resolve connectivity.

4. Ask for one response#

zeroclaw --config-dir ./qwen-test agent -a assistant \
  -m "Reply with the single word Ready. Do not use tools."

Our CLI returned Ready.. To test the browser too, run the gateway with the same configuration:

zeroclaw --config-dir ./qwen-test gateway start --port 42619

Open http://127.0.0.1:42619, pair using your own instance's code, then choose Open chat on assistant. Preserve the release's web/dist assets; see dashboard troubleshooting if the page does not load.

Running gateway dashboard showing the configured assistant

Open full-size dashboard screenshot. Enter the prompt and press Send:

Prompt entered in the real ZeroClaw chat before sending

Open full-size prompt screenshot. The completed conversation returned Ready:

Real Qwen3.5 response in the ZeroClaw dashboard

Open full-size response screenshot. Output punctuation can vary; a short reply proves basic connectivity, not general model quality.

5. Measure memory instead of guessing a minimum#

While the model remains loaded, run:

ollama ps
curl -fsS http://127.0.0.1:11434/api/ps

Actual Ollama loaded-model response showing context length and memory values

Open full-size memory screenshot.

Observation from this runMeasured value
HostApple M3, 8 GiB unified memory
Downloaded model data1.32 GB decimal, Q8_0 GGUF
Requested context4,096 tokens
ollama ps587 MB, 100% GPU, context 4096
/api/ps size and size_vram587,517,130 bytes each
Ollama runner process RSS snapshot1,396,192 KiB, about 1.33 GiB
ZeroClaw CLI maximum RSS34,209,792 bytes, about 32.6 MiB
First measured CLI turn, including model loading10.30 seconds wall time

These metrics describe different layers and must not be added together as independent allocations. Ollama's reported model allocation is not total system RAM, and an RSS snapshot is not peak inference memory. The 8 GiB host ran this small test; we did not establish a minimum RAM requirement. Keep headroom for the OS, browser, model server, KV cache and other workloads. Larger models, longer context and concurrent requests require a new measurement. The catalog's advertised maximum context is not what we tested.

Choose a larger model only after evaluating your actual task and measuring memory pressure. A 0.8B smoke test is useful for wiring checks; it does not establish dependable coding, planning, vision or autonomous tool use. For a machine that cannot accommodate the desired model, compare hosted providers.

6. Fix “selected model is unavailable”#

We sent the first request before the model pull had finished. ZeroClaw displayed:

The selected model is unavailable. Check the configured model name.

Actual missing-model error before the download completed

Open full-size error screenshot.

Wait for ollama pull qwen3.5:0.8b to finish, confirm the exact tag in ollama list, then retry the same prompt. Our retry returned Ready:

Same conversation showing the failure followed by a successful response

Open full-size recovery screenshot. In this release the earlier error banner remained until the page was reloaded; the new assistant message is the recovery evidence.

For a connection error, confirm the URI and test /api/tags. Our deliberately wrong URI http://127.0.0.1:1 produced a local-server-unavailable error; restoring port 11434 restored responses. See the verified troubleshooting fixes.

Sources: ZeroClaw v0.8.5 schema, Ollama Qwen3.5, and Ollama running-model API. Tests and source checks dated 16 October 2026.