OPEN LINK ↗
// 55d agoMODEL RELEASE
Jackrong drops Qwopus3.5-27B-v3 for steadier tool use
Qwopus3.5-27B-v3 is a Hugging Face model release from Jackrong built on Qwen3.5-27B, with a focus on reasoning stability, tool-calling reliability, and cleaner intermediate reasoning. The discussion around it is mostly about getting it to run under vLLM: the base Qwen3.5 family is supported by vLLM, but this specific checkpoint appears to need tokenizer/config adjustments, and the community report shows a startup error tied to the tokenizer class plus a follow-up warning about incorrect regex handling.
// ANALYSIS
Hot take: this looks promising as an agent-oriented 27B reasoning model, but it is not a smooth out-of-the-box vLLM experience yet.
- –The model is explicitly positioned for structured reasoning and tool use, not just raw chat quality.
- –vLLM supports Qwen3.5 in general, so the architecture is not the blocker.
- –The reported failure is a repo-specific integration issue: `TokenizersBackend` is not recognized by vLLM 0.18.1.
- –A workaround mentioned in the discussion is changing `tokenizer_class` to `Qwen2Tokenizer` and enabling `fix_mistral_regex`, but that still did not restore tool calling for the reporter.
- –Best read: this is a useful model for local agent workflows, but serving it in vLLM likely needs template/tokenizer cleanup rather than a simple `vllm serve` command.
// TAGS
qwen3.5qwenreasoningtool-callingvllmhuggingfacelocal-llm27b
DISCOVERED
55d ago
2026-04-03
PUBLISHED
55d ago
2026-04-03
RELEVANCE
8/ 10
AUTHOR
beginor