Problemsby IbisQwen upgrades break working setups
Outages and change3 posts from 3 people3d active+11 post in the last 7 days, 0 the 7 days before (steady)
Posts per day
The posts behind it
3, newest firstqwen3 parser: thinking starts with the newline after an explicit <think> because the tag and the newline arrive in different chunks### What is the issue? With `PARSER qwen3-thinking`, when the model writes its own ` ` tag (its trained format is ` \n...`), the thinking field of every response begins with `\n`. The content is clean; the thinking is not. `Qwen3Parser` strips one explicit leading ` ` and trims...ollama/ollama · mertkayacs1 repliesOpen ↗MODEL_FOR_CAUSAL_LM_MAPPING[Qwen3_5Config] points at the flat class, but all Qwen3.5/3.8 checkpoints are conditional-generation### System Info - transformers version: 5.14.1 (also present on main) - Platform: Linux x86_64 - Python version: 3.13 ### Who can help? @zucchini-nlp (Qwen3.5 area per related threads) ### Problem `MODEL_FOR_CAUSAL_LM_MAPPING[Qwen3_5Config]` resolves to `Qwen3_5ForCausalLM`...huggingface/transformers · micdoh1 repliesOpen ↗Qwen2Tokenizer overwrites pre_tokenizer from tokenizer.json. Wrong token IDs for all Qwen3.6 / Qwen3.8 models (v4→v5 regression)### System Info - transformers version: 5.17.0 (behaves correctly on 4.57.6) - Platform: Linux aarch64, Python 3.12 - tokenizers version: 0.23.2 - huggingface_hub version: 1.8.0 - PyTorch version: not installed (tokenizer-only reproduction) - Using distributed or parallel set-up...huggingface/transformers · simonepstein1 repliesOpen ↗
Companies and products named
About this problem
Evidence
3 posts from 3 people in 2 places, about 1 a week over 26 days. Mostly on huggingface/transformers, ollama/ollama. Tools named alongside: Hugging Face, PyTorch, vLLM, Ollama.
Frustration Frustration 0 of 3· Seen on GitHub issues
How it was grouped
Posts that state a pain and match the "breaking-changes" rule about Qwen. First post Sep 23, 2026, latest Oct 18, 2026. Corroborated: 3 or more posts from 2 or more people or places. Method
The same problem elsewhere
Upgrades and new releases break working setupsGitHub upgrades break working setupsClaude Code upgrades break working setupsCodex upgrades break working setupsOpen WebUI upgrades break working setupsClaude upgrades break working setupsAnthropic upgrades break working setupsClickHouse upgrades break working setupsKubernetes upgrades break working setupsChatGPT upgrades break working setupsOllama upgrades break working setupsOpenAI upgrades break working setups
Other Qwen problems
History
unverified to corroborated
fading to steady
steady to fading
Qwen upgrades break working setups (2 posts)