qwen3.6-35b-a3b and qwen3.6-27b infinite loop

[undici77]

Hi, I tried many versions of these models (oQ/4bit/8bit) but issue is always that same, after some interaction model enter in infinite loop!
Same models (I shared same files) with LM STUDIO work fine!

[zviratko]

What are you using it with? I found that having no system prompt (default in the built-in chat) makes the models get weird. Just putting something like "Be useful and concise" is enough.

I am trying qwen3.6-35b-a3b with hermes agent and it gets loopy on tool calls - might be related to the preserve_thinking issue (there's also one for hermes agent and I saw something for openclaw as well)... or maybe they just broke it, looks like local models are second class citizens to huge cloud ones.

[TipKnuckle]

I have had the same experience with looping with the new omlx 3.7 release. I switched back to omlx 3.6 and have not had the looping issue since. But I have not investigated thoroughly, so I can't say for certain that's the cause.

[fry69]

I switched back to omlx 3.6 and have not had the looping issue since.

Same here, I also switched back to 0.3.6 due to this problem. 0.3.6 is rock-solid for me, without endless loops.

See also issue #911

[jundot]

@undici77 @zviratko @TipKnuckle @fry69

Just pushed v0.3.8.dev1 for testing. Previously I was using mlx-lm as the decoding engine for VLMs because mlx-vlm didn't support continuous batching. Now that mlx-vlm has native continuous batching, I switched the VLM decoding engine over to mlx-vlm, which improved behavior across all Qwen models.

Could you give it a try and see if the looping issue is still happening? Any feedback is helpful since this is a dev build.

Release: https://github.com/jundot/omlx/releases/tag/v0.3.8.dev1

[fry69]

Could you give it a try and see if the looping issue is still happening?

Thanks for the quick reaction/release. Looks good so far 👍

I am running 0.3.8.dev1 now and have generated 3 moderately long/complex prompt/request pairs with qwen3.6-35b-a3b, which in the past ended up in endless loops, but work with this version. I have not tried qwen3.6-27b.

If those loops appear again, I'll report.

[molenick]

I downloaded 0.3.8-dev1 and still got looping issues with qwen3.6-27B. I also tried completing the work with qwen3-coder-next and ran into similar looping issues with that model as well.

Went back to 3.6 like the others have mentioned and it's plodding along with no looping issues yet

[jundot]

Thanks for the report, @molenick. I'll investigate more the looping issue with qwen3.6-27B on 0.3.8-dev1 and take a look at qwen3-coder-next as well. Will follow up once I have more info.

[lobsterbuko]

I can confirm, similar to @molenick, tremendous performance improvement by moving back down to 3.6, looping no longer occurs during ambitious prompts. Experimented with Qwen3.6-35, Gemma4-31, Qwen3.6-27, MiniMax-M2.7 (but that was an oQ2 quant, so almost lobotomized anyway!). Tried on 0.3.7, 0.3.8-dev, and 0.3.6: clear looping problems begin at version 0.3.7 and are not addressed so far in 0.3.8-dev1

[VictorLowther]

Yeah, I am seeing the same looping issues with qwen3-coder-next and qwen3.6-35b-a3b -- I saw them in 0.3.7 (installed via home-brew) and v0.3.8-dev1 (installed via dmg). Retrying with 0.3.6 now.

[kuanjames]

Same looping issue with Qwen3.5-122B-A10B on 0.3.8dev1

[beamivalice]

Can confirm that Qwen 3.6 - 27b produce a loop with only WARNING WARNING ... in the output (0.3.7)

[asxzy]

Same issue here. 0.3.6 works just fine. LMStudio with mlx-vlm backend also works works. Only 0.3.7 has the infinite loop issue. Tried enable/disable dflash with no luck.

[undici77]

What are you using it with? I found that having no system prompt (default in the built-in chat) makes the models get weird. Just putting something like "Be useful and concise" is enough.

I am trying qwen3.6-35b-a3b with hermes agent and it gets loopy on tool calls - might be related to the preserve_thinking issue (there's also one for hermes agent and I saw something for openclaw as well)... or maybe they just broke it, looks like local models are second class citizens to huge cloud ones.

I'm using it with qwen-code-cli, the official qwen tool for coding

[Jasshl]

I can confirm the same behavior. Using qwen3.6-35b-a3b with hermes agent and it was constantly looping on one task I tried. I tried the same task like 6 times and it was always looping. It was an issue both on 0.3.7 and 0.3.8-dev1. Downgrading to 0.3.6 fixed the issue.