Search results

Qwen3.5-4B CEval accuracy drops ~4.1% when using SGLang and VLLM

evalQuestions related to evaluation resultsQuestions related to evaluation results

Status: Open.

#137 In QwenLM/Qwen3.6;

· AuFlow opened on Apr 16, 2026

incorrect processing of a line of code

qwen.aiProblems with Qwen.AIProblems with Qwen.AI

Status: Open.

#134 In QwenLM/Qwen3.6;

· EBOC834 opened on Apr 11, 2026

chat template emits empty historical <think> blocks, causing prompt drift and cache invalidation

Status: Open.

#131 In QwenLM/Qwen3.6;

· latent-variable opened on Apr 9, 2026

Unable to reproduce HLE benchmark score with tool-use (Qwen3.5-122B-A10B)

evalQuestions related to evaluation resultsQuestions related to evaluation results

Status: Open.

#126 In QwenLM/Qwen3.6;

· CloudyDory opened on Apr 6, 2026

Qwen3.5-9B / 35B-A3B produce XML-style tool calls instead of JSON under SGLang tool-call parser

badcaseBadcases from the modelBadcases from the model

Status: Open.

#125 In QwenLM/Qwen3.6;

· BlackJack0083 opened on Apr 5, 2026

使用最新的VLLM部署qwen3.5-9B，输出的内容都在think中，lmdeploy也是。

Status: Open.

#122 In QwenLM/Qwen3.6;

· LIUKAI0815 opened on Apr 3, 2026

Error occurred when qwen3.5-27b invoking tool

Status: Open.

#117 In QwenLM/Qwen3.6;

· hxujal opened on Mar 31, 2026

Cannot disable Qwen3.5 reasoning output via Dify; reasoning process is in English by default

Status: Open.

#116 In QwenLM/Qwen3.6;

· xiaolvtongxue-zt opened on Mar 31, 2026

Qwen3.5-27B generates gibberish/repetitive loops with long context on vllm 0.17.0

Status: Open.

#115 In QwenLM/Qwen3.6;

· haomiemie opened on Mar 31, 2026

Proposal: new chat_template_arg enable_history_reasoning for reusing prompt cache among querys within Agents .

enhancementNew feature or requestNew feature or request

Status: Open.

#112 In QwenLM/Qwen3.6;

· Abioy opened on Mar 24, 2026

Unable to reproduce SWE-Bench Verified scores due to missing agent framework and max_turns configuration

evalQuestions related to evaluation resultsQuestions related to evaluation results

Status: Open.

#110 In QwenLM/Qwen3.6;

· Hellozaq opened on Mar 23, 2026

SGlang run error: ValueError: Speculative decoding for Qwen3_5MoeForConditionalGeneration is not compatible with radix cache when using --mamba-scheduler-strategy no_buffer.

Status: Open.

#108 In QwenLM/Qwen3.6;

· yonghenglh6 opened on Mar 23, 2026