DOAMOD

https://i.redd.it/o9ipr4utix4h1.gif

https://preview.redd.it/euxlf9tz4x4h1.png?width=2169&format=png&auto=webp&s=460276d887c6dd11c93773fbdf5db5f28db74f7e qwen 3.6 27b

No way is this bigger than mm2.7, if so this is very disappointing, I've been testing it and it just isn't.

like a teenager?

Yes, maybe now 3.7

Qwen are you crazy?

https://i.redd.it/2sg41k1g2q1h1.gif MTP@110 toks

https://i.redd.it/pzqkfz5kyp1h1.gif Q4XL

Q8 Run 3 https://preview.redd.it/h1dhz50dbyzg1.png?width=970&format=png&auto=webp&s=8b5ed52bdc94d87c61a2e92c7e1e363377a3eae2

Q8 Run 2 https://preview.redd.it/plnjf42layzg1.png?width=1250&format=png&auto=webp&s=14dd6454f697479c206ab3a7d49ac9c46f583bcd

Q8 Run 1 https://preview.redd.it/a1p251do9yzg1.png?width=1332&format=png&auto=webp&s=6c1b5fda53ca701873288bf1227a8da24349e915

Q6XL Run3 https://preview.redd.it/yslhp6dn7yzg1.png?width=1209&format=png&auto=webp&s=95036b3b28fd8f0c3e4244645cfcccee67ce1169

Q6XL Run 2 https://preview.redd.it/2qftt5bw6yzg1.png?width=1295&format=png&auto=webp&s=cd77418c3068865e5c28b4e65634204bbb59e0f4

Q6XL Run1 https://preview.redd.it/r3m3o2nj5yzg1.png?width=2451&format=png&auto=webp&s=d38ab34e3edbf7bfaf23e1312f77d6a10bdbb73f

Q5XL Run3 runerrorsfix https://preview.redd.it/8cmziknz2yzg1.png?width=1076&format=png&auto=webp&s=e4045ec6d9246124ab5957cbcb761b57a6e55b53

Q5XL Run3 https://preview.redd.it/o2y7zk292yzg1.png?width=2702&format=png&auto=webp&s=b232a185fd0b70d8fc937c1ffc116b61e05f620e

Q5XL Run2 https://preview.redd.it/70zlllmz1yzg1.png?width=1165&format=png&auto=webp&s=2b789beb9aaecf5538125b3f3fb9eca45ea93f04

Q5XL RUN1 Autothinkfix1 https://preview.redd.it/7nw9492p1yzg1.png?width=1161&format=png&auto=webp&s=17e418157b996797a6e69d87ffec1146e7babe44

Q5XL RUN1 (not final) https://preview.redd.it/4woxo3of1yzg1.png?width=1163&format=png&auto=webp&s=8b798252685aeced2ada8b45d1f507d6a2284901

Q4XL Run3 https://preview.redd.it/lal389nv0yzg1.png?width=1263&format=png&auto=webp&s=97339f4bc279c64250531a4fcd84bd6e2f771121

Q4XL Run2 https://preview.redd.it/wwjgz1bq0yzg1.png?width=1521&format=png&auto=webp&s=d9d71313443421d24d3cb3ae777cbc53f2a4a704

Q4XL Run1 autothinkfix final https://preview.redd.it/3d4bevue0yzg1.png?width=1255&format=png&auto=webp&s=2cc3326ee33cdf688eed8cdf8f5eaaaf2ec464e4

My test 3.6 27B 3RUNS for Q4XL-Q5XL-Q6XL-Q8 Q4XL Run1 (not final run) https://preview.redd.it/2pe6gtdxzxzg1.png?width=1260&format=png&auto=webp&s=08320288eef7a9a0d0f185dd04da232dca302a4d

I haven't tried Llamacpp MTP yet, but I did try MTP in VLLM on Windows on my 5090, and it was a bit disappointing. The memory consumption when exposing the small model doesn't compensate at all for the significant loss of context window. Perhaps in some specific cases for MoEs it could be useful; I think that's the interesting point. But for Dense, I don't see a benefit in my use case. I'll try Llamacpp, though.

That's very human

Yes, same for me, some calls fails with 3.6 27/A3

Yeah

The 27 is actually quite a bit better. I've been working with it for several hours, and the difference is noticeable in something you realize very quickly: the 27 doesn't have to exert much effort, it works well and makes almost no mistakes, while the 3.6-A3 has to struggle, consuming an overwhelming amount of context and making many more simple errors. They're both truly incredible, and I love them, but clearly the a3 reaches its level through a lot of effort, and that's no small feat.

Very first test much better less thinking and less syntax errors over 3.6-A3

It's crazy

The king

This question appears to have been success in 3.6-A3 **3.6 35-A3**(only for this, test 4-5 times on q4-q8 all success): That's a fun question with a clever twist — you're washing a car, and the car needs to be *at* the car wash! 😄 Here's the thing: you have to drive . Walking won't work because your car is still at home. You'd end up walking 50 meters to the car wash and then having to explain to someone why you showed up alone without a vehicle. That said, the distance itself doesn't really matter — whether it's 50 meters or 5 kilometers, driving is the only practical option since the object being serviced (the car) needs to travel with you. The real decision here isn't walk vs. drive — it's whether you want to wash it yourself or let someone else do it at a professional car wash! If you're already driving there, might as well enjoy the ride and get it properly cleaned. 🚗✨ **Gemma 4 26-A4**: Honestly? Walk. At 50 meters, you are looking at a distance that takes about 30 to 45 seconds to cover on foot. If you decide to drive: 1. The Effort-to-Reward Ratio is terrible: You'll spend more time starting the engine, shifting into gear, and maneuvering out of your driveway than you would simply walking. 2. Fuel & Wear: You'd be burning fuel and adding unnecessary mileage to the very vehicle you are trying to clean. 3. The "Car Wash Paradox": Driving a dirty car to a car wash is the standard procedure, but driving it only 50 meters feels like using a sledgehammer to crack a nut. Verdict: Save the engine the trouble and just enjoy a quick stroll. Plus, you'll be ready to head straight into the wash without having to find a parking spot again! **Gemma4 E2**: Hello. Given that the car wash is only 50 meters away, the decision is quite simple, but it depends on your preferences and what you are doing at that moment. Here is a breakdown so you can decide: # Option 1: Walking 🚶‍♂️ * **Advantages:** It is fast for such a short distance and is excellent if you want to get a little exercise or if you have things to carry. * **Disadvantages:** If it is very hot or raining, walking might be uncomfortable. # Option 2: Driving 🚗 * **Advantages:** It is more comfortable if you are carrying many things or if you don't want to exert yourself. * **Disadvantages:** It might require finding parking near the car wash and could be a bit more cumbersome than walking 50 meters. **My Recommendation:** For a distance of only 50 meters, **walking is probably the quickest and simplest option** , unless you have a specific reason to use the car (like carrying a lot of luggage). What do you prefer to do right now: move quickly or be more comfortable? 3.5 A3: # My Recommendation: Walk to the car wash. At 50 meters: |Method|Estimated Time|Fuel Used|Logistics| |:-|:-|:-|:-| |🚶 Walk|\~1 minute|0%|Leave car where it is| |🚗 Drive|\~2-3 minutes|Used|Unnecessary movement| # Conclusion: Walking is clearly the more efficient option. It's a perfect walking distance, and driving would just waste your time, burn fuel, and potentially stress you out with maneuvering and parking for such a short task.

They could be a jewel

It's strange, but the Q5 was also working worse for me. Have you tried the Q6?

This model consumes a lot of context in its reasoning, and if it becomes too complex, it can get lost and descend into chaos. Yesterday, I encountered a situation where it became obsessed with a closure it couldn't find and started going around in circles, trying suboptimal solutions. We need to be careful because this is the biggest problem I'm seeing with this model; basically, it can get lost due to its own reasoning. In my tests, it's on par with model 27, but it requires a much greater mental effort, which consumes large windows of context. It's very context-hungry, and this is its weak point because all its speed advantage is lost in its reasoning for solving problems, in my test by solving code problems. It was about 15 seconds slower than model 27b, despite being much faster.

New fresh one shot and 2 fixes(easy first start fix and long cars form), fps, pedestrians are very well https://i.redd.it/31jr60q6pzvg1.gif

In my tests, 3.6 A3 rivals 3.5 27, but it's about 15 seconds slower by problem. for me its crazy, I cant wait for 27b...

Those of us who actually use the model and aren't just talking nonsense, said so from day one, and people saying this is just benchmarxx.

By the way, the game has collisions, and the cars explode with fire, particles, and smoke.

https://preview.redd.it/sdre0le3tyvg1.png?width=1542&format=png&auto=webp&s=ba7f84980dbd2144fbf657695090ff8e3056c674 haha, after fighting with floating buildings, its crazy this model.

Q3.6 is insane, it did real-time day/night weather with rendered shadows, the entire weapon system, the STARS search system, this is just one shot and a fix for the loading screen that wouldn't start, now we would have to fix the floating buildings, but it's really impressive for such a small model. https://i.redd.it/7r42se3ebyvg1.gif

Thanks, I have a question: are the mmproj versions different between 3.5 and 3.6? Do they have any improvements or are they essentially the same? I'm not sure whether to update them; it's a bit chaotic with so many.

You're the first person I've seen who thinks the same as me. It seems surprising to me all I read about people defending Gemma4, when it's so lazy. It's exactly the definition I've been thinking about for days every time I use it: it doesn't want to do anything. It even admitted it to me, saying it's more of a conversational model. What a surprise, it's a chat model, yes, very intelligent, and writes very well, but it's not your coworker. You know what else it told me? It told me to go to YouTube and search for the information or on Google. OMG, never in my life has a model told me to look up the information myself, hahaha this was incredibly fun.

):

https://i.redd.it/zx39zz4dqlvg1.gif

I'm testing it out and it's thinking a lot, but it seems very intelligent. I think I'm going to like it. I'm really looking forward to seeing the 27b and what it can do.

Gemma 4 is dead?

I'm using 2.7 and I'm surprised at how well it's working, looking at that graph, I also agree, it seems surprisingly good for what one would expect.

He spends the day wandering around

still broken