Qwen3.8-FlashNext vs Qwen3.8-27B tool calling issues
retained text
I've been using the 27B series for a while now. It's so good, I don't even use a custom chat template and **barely ever saw tool call failures**. My latest model & engine of choice were `ukisai/Swift-1.5-Qwen3.8-27b-NVFP4` running on `NInfer`. Recently I began experimenting with QFN, more specifically `ukisai/Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-GGUF:IQ3_S` running on `Basalt`. The model seems pretty good, but I see quite a lot of tool call failures. Sometimes I see the model chaining up to four or five failed tool calls until it *finally* gets it right. Using the exact same harness & setup for both (OpenCode). Tool calls could be file edits, playwright commands, or anything really. I'm unsure if this is a quantization issue (IQ3_S), a model issue (QFN), or an engine issue (Basalt). Has anyone experienced tool calling issues with QFN? Because it sure was **not** an issue with Q3.8-27B on any quant (4bpw+) and engine (llama.cpp, vLLM, exllamav3, NInfer) I tried. I'm planning some runs of BFCL v3 or v4 with different configurations to see if I can collect better data on tool calling. I'd appreciate if anyone would share their findings on tool calling success rates with QFN, and if there is anything that can be done to improve it. **EDIT:** Thanks for the replies. Most shots fired were at the quantization, some at the engine, none at the model, so that's the order of factors I'll tinker with.
document fetched 2026-10-21 00:41 · canonical https://reddit.com/r/LocalLLaMA/comments/1xaos52/qwen38flashnext_vs_qwen3827b_tool_calling_issues