Issues

• [Feature Request] 请求适配 LiquidAI LFM3-VL-2B 多模态模型

  alibaba/MNN#5104 · Filter by author Alivecat opened on Nov 4, 2026
  

• MNN 3.6.3 Roadmap

  alibaba/MNN#5101 · Filter by author wangzhaode opened on Nov 4, 2026
  

• mnn 3.6.2 Qwen3.5-2B-Dflash 推理报错：CPULinearAttention: cannot erase part of the 34-token history; reset all of it instead

  alibaba/MNN#5093 · Filter by author wangpx5 opened on Nov 3, 2026
  

• [LLM] set_config updates sampling values without refreshing the existing sampler

  alibaba/MNN#5078 · Filter by author Torry2022 opened on Oct 31, 2026
  

• [CPU] Later runtimes are capped by the first thread pool for the same CPU mask

  alibaba/MNN#5077 · Filter by author Torry2022 opened on Oct 31, 2026
  

• [Bug] Qwen3-TTS 0.6B: generation never emits EOS and decodes to digital silence for some texts (int4/int8/int3/fp16)

  alibaba/MNN#5062 · Filter by author Pokonliu opened on Oct 29, 2026
  2

  

• [Bug][CPU] Arm82 fp16 AvgPooling 使用 fp16 累加器 —— 大窗口平均池化输出系统性衰减至 fp32 基准的约 25%

  alibaba/MNN#5043 · Filter by author harhar539 opened on Oct 26, 2026
  

• [LLM] prompt_cache 复用后的条件 tokens 与完整模板不同，是否为预期行为？

  alibaba/MNN#5040 · Filter by author kerry0520 opened on Oct 23, 2026
  

• [CPU] Dilated convolutions on small feature maps spend substantial work on padding

  alibaba/MNN#5018 · Filter by author patrickkabwe opened on Oct 20, 2026
  

• [CPU] Dilated convolution computes taps that always read padding

  alibaba/MNN#4999 · Filter by author patrickkabwe opened on Oct 19, 2026

• 请求适配 Florence-2-base 模型

  alibaba/MNN#4995 · Filter by author Rye426 opened on Oct 19, 2026
  1

  

• [CPU] FP32 GELU_STANDARD lacks NEON acceleration on ARM64

  alibaba/MNN#4994 · Filter by author patrickkabwe opened on Oct 19, 2026
  

• [Converter] 动态形状 fp32 转换产出与输入无关的常数输出（MNN 3.6.1）

  alibaba/MNN#4986 · Filter by author modacker opened on Oct 18, 2026
  

• Question: Getting started with MNN on Qualcomm Android devices

  type:question
  alibaba/MNN#4981 · Filter by author zehnderdev opened on Oct 16, 2026
  1

  

• [Feature Request] 请求适配 Qualcomm Hexagon NPU 后端

  alibaba/MNN#4980 · Filter by author 0Sakura721 opened on Oct 16, 2026
  

• [CPU] Caller affinity is not restored after runSession

  alibaba/MNN#4979 · Filter by author patrickkabwe opened on Oct 16, 2026
  1

• [QNN] StridedSlice rejects NC4HW4 input after convolution

  alibaba/MNN#4978 · Filter by author patrickkabwe opened on Oct 16, 2026

• Generation error in the Sana-Edit model editor

  alibaba/MNN#4966 · Filter by author alexdsh opened on Oct 13, 2026

• qwen3-asr-0.6b类似的语言模型，后续是否会在modelscope上出官方的转换版本

  alibaba/MNN#4961 · Filter by author vklonghml opened on Oct 13, 2026
  

• 加载文生图模型崩溃闪退

  alibaba/MNN#4958 · Filter by author nbbjs opened on Oct 12, 2026
  1

  

• [Metal] Potential mmap weight-cache corruption when reusing stale cache or switching cache modes

  backend:metaltype:bug
  alibaba/MNN#4939 · Filter by author kerry0520 opened on Oct 8, 2026
  1

• ONNX Softmax before opset 13 uses the wrong omitted-axis semantics

  module:convertertype:bug
  alibaba/MNN#4935 · Filter by author hulkbig opened on Oct 5, 2026

• [CPU] DynamicQuantizeLinear half-tie rounding changes converted MatMulInteger output

  module:quantizationtype:bug
  alibaba/MNN#4924 · Filter by author hulkbig opened on Oct 2, 2026
  1

• share unwrapping cl_mem of Interpreter for Zero-Copy

  backend:opencltype:feature
  alibaba/MNN#4915 · Filter by author fatalfeel opened on Sep 29, 2026
  1

  

• 创建推理会话时模型张量名进入默认哈希表的潜在哈希碰撞 CPU 放大

  status:needs-infotype:performance
  alibaba/MNN#4901 · Filter by author K-ANOY opened on Sep 22, 2026