That's a cool write up! I suspect a lot of this will have to do with quantization, how much quality is getting lost, and how each model behaves under different quants.

Thank you for the reply Yes moe models are much faster and usually proportionately dumber… I run some tests and the 26b results are unexpectedly good, hence was the question .

https://www.glukhov.org/ai-devtools/opencode/llms-comparison/

Apples and oranges- don’t care how they are organised internally when they deliver quality output. Faster better, but if they don’t deliver - who cares.

You're comparing apples and oranges. Qwen3.5 27B is a dense model. Gemma4 26B is a mixture of experts model with 4B parameters activated at once. The equivalent would be Gemma4 31B, which is the Gemma4 dense model.

Both dense models are EXTREMELY good. From my testing, they can code and work agenticly with similar performance as a cutting edge model (Gemini, ChatGPT, Claude Sonnet) from 4-5 months before their release. Usually, Gemma models are better at prose and the Qwen model has scores a little bit better on coding and logic tests. These models being dense, they require more computation and memory bandwidth than mixture of experts (moe) models, which means they're slower or more expensive to run.

If you purely are comparing the models you originally listed, the Qwen model will crush Gemma4 26B but it will run at a quarter of the speed. :)

You mean 3.6? I tried it with opencode, with the model hosted on opencode zen - very smart and fast… Hopefully they will release its smaller variant

I've been impressed with Qwen 3.5 122B - can't speak for the lower size one though. Very good for agentic, on par with Sonnet IMO. Excited for the 3.6* release.

I don’t really know good OCR outside of LLMs. I always find it a pain to setup and limited… sorry

Are there any other open-source options with great performance that don’t use Ollama?

Nowadays you can get really good small Qwen3.5 models, but it might be overkill for the task?

I just can’t take Mozilla seriously anymore for pretty much anything.

This is a review: I have the Orange PI AI studio thing, 96gb ram and two NPUs, using gpustack, the best I've gotten is qwen3-30b-a3b with around a 100k context window - runs at 25-28tps

There are some weird quirks with it, I've mostly been constrained to using MindIE as a backend which doesn't support all models. I've tried vLLM but it tends to crash and complain it can't talk to both NPUs.

Overall I'm pretty happy with it though

Holy shit this is cool. No price or reviews yet though, but will be keeping my eyes out. Sounds great for my k8s cluster