Are any of you experiencing Tool Call Issues for the following models?

[BradKML]

I would like to note how, at a quick glance at social media, there are a lot of weird phenomena popping up:

• GLM 4.5 is surprisingly good for flexible tool calls
• Qwen3 instruct models are more compliant with new tool cal standards, while Qwen3-Coder models are made for their own tools
• DeepSeek v3.1 went the opposite direction of Qwen3, avoiding the formatting issues of the original R1 and v3
• Kimi K2 and/or Kimi Dev are being noted to be some of the worst tool-calling LLMs, but they are not overtrained to fit custom formats
• It is likely Qwen3 is inspired by both GPT-OSS models with their Harmony format, AND Anthropic's API formats
• Old names like Microsoft's Phi-4, Meta's Llama-4, and Mistral, are now behind the curve

Some cross-reference notes: UfoMiao/zcf#31 (reply in thread)
https://gist.github.com/WolframRavenwolf/0ee85a65b10e1a442e4bf65f848d6b01?permalink_comment_id=5733908#gistcomment-5733908

[uinstinct]

Is this with native tool calling or with the system message tools?

[BradKML]

Let's cover both, but assume most services have native tool calling like OpenRouter and Ollama and the rest.

[BradKML]

Also @uinstinct please spread the word if possible, need to check the worst offenders when it comes to LLM tool calls vs BFCL Kilo-Org/kilocode#2107