Needle Tops Qwen3 In CPU Tool-Calling

This is a useful reminder that “tool-calling model” and “chat model that can call tools” are not the same product class. Needle looks like a real dispatcher for fixed tool palettes; Qwen3 looks like the safer generalist once you need language robustness or broader conversational behavior.

• –Needle wins on tool_match overall and latency, but its main failure mode is wrong-tool routing, not argument quality.
• –Qwen3’s main failure mode is more brittle: it often never emits a tool call at all and falls back to prose.
• –The T3 implicit-intent tier is the real separator; Needle maps intent directly, while Qwen3 collapses when the tool name is not explicit.
• –The schema lesson matters: feeding Needle OpenAI-style JSON Schema tanked results until the author converted to Needle’s flat schema format.
• –The edge-case win for Qwen3 on Hindi/French shows this is not a universal small-model verdict, just a strong case for tiny specialist routers on CPUs.