Unfortunately I don’t have enough experience with Qwen 27B to immediately compare, but I do it’s Qwen 3.6 35B A3. It’s much slower obviously but it seems to be way more efficient with its thinking to the point that using it might actually be faster. I find Qwen and some others rehash the same things over and over when thinking without getting anywhere, in mg limited checks here Muse is much better.
I really like the 3.6 35B A3B for experiments, and it seems OK, but as you say, it spins round in thinking loops more than say the 26B Gemma 4 does. If Muse doesn't actually-wait itself as much it will be very interesting.
I am just downloading it to run my small tests.
In my experience it is almost never productively thinking past that point, just spinning in circles. I also reinject all of the thinking. And there are a few tells that it is getting stuck. I have an optional mode that takes the last few turns and tool calls and shoots it off to DSV4 with a prompt to basically understand where it is at and inject better thinking and or planning. It almost always gets it over relatively difficult humps, but some of the time I don't want things going remote. It might end up with 10-30 cents of DSV4 calls over a hours and the quality improvement is remarkable.
The other thing is I trick it into thinking a web_search tool is a web search but it really just asks DSV4 the prompt. DSV4 is a cheap filter to help prevent prompt injection lol. You can give it other models but DSV4 is my cheap-mode default.
edit: oh! My final 35B A3B tip -- use an extremely simple harness. Pi is good. Pi's default tools almost exactly match what Qwen says they tested the model with (likely meaning that tool set is also what they trained it with or something similar). So, in my experience bigger harnesses don't have a noticeable improve ment on tasks.
It had not occurred to me that you could nudge it to stop thinking with a proxy. Nice idea.
Will favourite your comment and come back to it.
ETA: Incidentally you've helped me put into words the difference between the way Muse Glimmer thinks to the way Qwen thinks. There is a clear sense of urgency in Glimmer's thinking traces.
I really do feel like it’s effective tok / s is way higher because it doesn’t waste them.
I don't really enjoy the way Qwen writes prose, and I find its thinking a bit exhausting, though it clearly writes very good code.
I like the neutral, clear way the Gemma models write, which I sometimes use to get myself a "getting started" document on something I want to understand; it also summarises well. It is neutral, sensible, un-showy. It writes in a way that is fairly close to what I would use for documentation. The 12B and 26B models are also very good for talking about art and photography. Analysing my own photographic work has helped me more than I expected it to.
This model, honestly, has made me smile. It also feels like it is more creative at a given temperature than Gemma. I am trying to motivate myself to do something quite open-ended so I asked it about what other people's considerations might be in my situation, and at the risk of anthropomorphising, the things it has come up with feel like the work of a more curious mind, somehow. More eclectic. I have enjoyed testing it and I really want to test it more, which might help me get over a motivation hump there, too.
(I am also exploring its hard-wired policies by asking it to analyse some studio art nude work I have done; it definitely thinks out loud about its policies in a way I have not seen Gemma do.)
I suspect it will help a lot with enabling preserve-reasoning, because the biggest apparent limitation of this model is the 128K context window.
Though the practical issue I am seeing on my M1 Max MBP is that performance suddenly drops off a cliff if I have DFlash enabled.
I added these arguments to my llama-server call, using the official GGUF release from Meta’s account on HF.
`--rope-scaling yarn --rope-scale 2 --yarn-orig-ctx 131072 --override-kv muse-glimmer.context_length=int:262144,dflash.context_length=int:262144`
(DFlash seems rather weird on my M1 Max; model performance suddenly drops off a cliff like maybe there is some memory management issue).
One thing I will say is that its thinking traces are really quite terse. It genuinely seems to spend many fewer tokens on reasoning. So that might help a bit.