[Bug] Qwen3.6 MoE output collapses into repetition/gibberish

Description

I was running the Qwen3.6 MoE model from the new release and it worked great! Then, after about 400 characters, the model suddenly started spitting out gibberish. This happened consistently across various prompts.

Environment

• FLM Version: v0.9.45
• OS: Windows 11 25H2
• CPU/NPU: AMD Ryzen AI 9 HX PRO 370 w/ Radeon 890M (2.00 GHz)
• System RAM: 64 Gigs, Task manager reports me getting up to about 24-26 so I should be clear of the 50% ram to NPU cutoff limit.
• VRAM / NPU: Allocation: 3.8 GB Dedicated (iGPU) / 29.8 GB Shared Pool

Reproduction / Context

The complete prompt was: "Write me a 1000 word story about a rabbit:"

Additionally, I was running Open WebUI, so several other requests are queued and created during the processing, but I'm pretty sure they are held by some middleware until the server is ready.

Something that might be relevant is that sometimes (but not always) I get this warning while loading the model into RAM:

[FLM]  Configuring NPU Power Mode to performance (flm default)
[FLM]  Using user-specified port: 8081
[FLM]  Qwen3.6 pre-resize image height to 720 pixels if larger than that
[FLM]  Loading model: [Path to model]
[info]  Still some invalid buffer, last try!

Logs

Here are the FLM logs: (This snippet includes server logs about network requests in case that information is useful).

[FLM]  Start prefill...
[FLM]  Total images: 0
[FLM]  Prefill chunk 1/1 with 666 tokens
[TCP]  Read 155 bytes from socket
================================================
[LOG]  Time stamp: 22:28:09 07:10:2026
[LOG]  Target: /v1/models
[LOG]  Version: 11
[LOG]  Keep-Alive: 1
================================================
[🔒 ]  TCP connection closed - Remote: 127.0.0.1:55605
[🔗 ]  TCP connection established - Remote: 127.0.0.1:55606
[TCP]  Read 2958 bytes from socket
================================================
[FLM]  Creating checkpoint at context length 666, general_token_idx 666
[FLM]  Start generating...

[FLM]  Model RAW Output:
The rain in the forest of Oakhaven was not a violent storm, but a gentle, persistent drizzle that turned the ground into a rich, dark sponge. For Barnaby, a rabbit with ears so long they often dragged in the dirt and eyes so wide they seemed to perpetually surprised, this was the best time of day. It was the time when the world slowed down, when the sharp edges of reality softened into a blur of greens and browns.

Barnaby was not like the other rabbits. While his brethren were content with their clovers and quick escapes from hawks, Barnaby was a philosopher of the under-leaf. He spent hours staring at the intricate patterns of moss on ancient stones, wondering if they were maps to hidden worlds or just random textures generated by nature’s whimsical hand. Today, however, he had a mission. A small, crumpled piece of paper lay in his paw, damp and fragile. It was a note from Mrs. Gable, the elderly woman who lived in the cottage at the edge of the clearing. *“Come visit,”* it said, with an arrow pointing toward her garden. Barnaby had never been invited before. Usually, he was merely tolerated, a nuisance to be chased away with a broom. But today, he was wanted.

He adjusted his scarf—a knitted thing of blue wool that Mrs. Gable had left on the windowsill weeks ago, which he had stolen for warmth. With a flick of his nose and a hop, Barnaby began his journey. The forest path,s quiet, not,, to,. not; with....—, more,,,,,,,,,,,,,.,,,,,,,,,, in, yet。,, in,,,,,,,,ty...:’s，, in—，arily,s,ly|,，—。lessly;part,…，arily, end,...arily。....。arily...arily Ray....arily...arily雷。arily...。arily.。。arily...arily雷。arily...arily...。arily雷arily
。arily
ChatCompletionChunk: {"id":"chatcmpl-32c81ecd371601aaa6966581","object":"chat.completion.chunk","created":1783744089,"model":"qwen3.6-moe:35b-a3b","system_fingerprint":"fp_c5033e3661633a03","choices":[{"index":0,"delta":{"content":null},"finish_reason":"stop"}],"usage":{"prompt_tokens":666,"completion_tokens":445,"total_tokens":1111,"active_kv_tokens":1111,"max_kv_token_capacity":131072,"kv_token_occupancy_rate_percentage":0.847625732421875,"load_duration":1e-06,"prefill_duration_ttft":28.31686656,"decoding_duration":59.575152,"prefill_speed_tps":23.519551451387706,"decoding_speed_tps":7.469557106627273}}
[🔒 ]  Closing TCP connection (streaming, non-keep-alive)
[🔗 ]  TCP connection established - Remote: 127.0.0.1:50829
[TCP]  Read 127 bytes from socket

Start by reproducing the 1000-word rabbit prompt with Qwen3.6 MoE on the reported Windows 11 AMD NPU setup, while checking the model-loading warning and the provided generation logs. Compare output before and after the roughly 400-character transition, including with queued Open WebUI requests; done means generation remains coherent instead of collapsing into repetition or gibberish.