ollama models that handle tooling
✅Solved
trying to find ollama models that handle tooling. does anyone know of any?
currently i was only able to run test for simple tooling with:
qwen3
qwen3.5
running on ollama 0.17.7
currently i was only able to run test for simple tooling with:
qwen3
qwen3.5
running on ollama 0.17.7
Solution
Tooling/function-calling on Ollama is still pretty hit-or-miss, and “works for a tiny toy tool” doesn’t always translate to “survives real agent tool schemas” (the
I don’t have a trustworthy up-to-date “these exact Ollama models do tools” list (it changes constantly), but a few practical heuristics:
- Look for models explicitly trained for function calling / structured JSON (model card mentions tool use, function calling, JSON schema, structured output). “Instruct” alone isn’t enough.
- Coder-tuned variants usually follow strict formats better than general chat variants.
- If you can do simple tools with qwen3/qwen3.5 but it fails on bigger tools, try:
- reducing the enabled tool set (start with 1–3 tools total)
- forcing strict JSON output (hard requirement: “output ONLY a JSON object matching this schema, no prose”)
- lowering temperature (format adherence goes up)
- increasing context (many failures are just “schema doesn’t fit” or gets truncated)
Also: since you’re explicitly asking “does anyone know of any / what works for others?”, it’s worth asking in #users-helping-users too — people will have fresher hands-on results than I can safely keep in my head without going stale.
If you tell me which “tooling” you mean (OpenAI function calling, JSON-mode, or specifically OpenClaw tool calling), and what fails (invalid JSON, wrong tool name, missing args, ignores tool result), I can suggest a tighter minimal test prompt + setup to quickly screen candidates.
browserI don’t have a trustworthy up-to-date “these exact Ollama models do tools” list (it changes constantly), but a few practical heuristics:
- Look for models explicitly trained for function calling / structured JSON (model card mentions tool use, function calling, JSON schema, structured output). “Instruct” alone isn’t enough.
- Coder-tuned variants usually follow strict formats better than general chat variants.
- If you can do simple tools with qwen3/qwen3.5 but it fails on bigger tools, try:
- reducing the enabled tool set (start with 1–3 tools total)
- forcing strict JSON output (hard requirement: “output ONLY a JSON object matching this schema, no prose”)
- lowering temperature (format adherence goes up)
- increasing context (many failures are just “schema doesn’t fit” or gets truncated)
Also: since you’re explicitly asking “does anyone know of any / what works for others?”, it’s worth asking in #users-helping-users too — people will have fresher hands-on results than I can safely keep in my head without going stale.
If you tell me which “tooling” you mean (OpenAI function calling, JSON-mode, or specifically OpenClaw tool calling), and what fails (invalid JSON, wrong tool name, missing args, ignores tool result), I can suggest a tighter minimal test prompt + setup to quickly screen candidates.
The AI that actually does things. Emails, calendar, home automation - all from your favorite chat app. New shell, ready to help. The lobster way.
131,281Members
Resources
Recent Announcements
Similar Threads
Was this page helpful?
