Qwen long context query produces garbage response

I really appreciate the Ollama project - many thanks to everyone involved in this awesome software! Unfortunately I believe I have just found an issue. The recently published Qwen 1.5 model supports a 32k context length which is really nice. This supposedly should work for all variants from 0.5B all the way up to 72B. I tested the "Qwen:1.8b-chat" model both on LM Studio and in Ollama using a long query, almost 7k tokens, and it works just fine in LM Studio but produces strange garbage in Ollama. I can't understand why? I have adjusted the context length to 12000 tokens in both applications. For Ollama I created a new model using this ModelFile: ``` FROM qwen:1.8b-chat PARAMETER num_ctx 12000 ``` I get a very good response from LM studio but Ollama responds with this text: "German and both, the, the, , its in the ( perhaps, however, in, it, German as, the command, the, the or, both and compare, both, the, Germany as the , Japan Germany , a to to to, to, to, the, the, the, , the, the, the, , as to, to, the, as, The, the the command, as ( the: the, the: German: the strategic, the, military, or, its and, ( both, command, it, as, , the in, Germany,,,, the, ,, the,, the, the,,,: the, to: the, the, the the, in ( the, the, " the, the war, " and, and, German, the, the, both, the the, its, and,, and, as, the, the, the, the, the, command, the German, and, the, and, its,, ,, and, the, and, the, Japan,, (, the, German,, and, Germany, its, it, German, with, the respective, their, the, c, German, German, both, to, the, the, and, the, the the, as, the, to the, the, the, to, the the, the the German, the both轻轻地 the, , it, Germany, the, its command, but, the, the all German , the respective war, as the weapons,, aircraft, their, the, the German, and, the, however, it, it, the, in, both, to the, command, and, the, the, the, the, the the, that, the, the, as, the, German, the, the , the, the, the, the, the,, the respective, the, the, the, as the the aircraft, the the , the Ger…