google/gemma-4-31B-it

Image-Text-to-Text • 33B • Updated 11 days ago • 4.47M • • 2.25k

Note Excellent official model, after a very boring 2025 year, 2026 is shaping to be awesome. It's surprisingly compliant for a corporate model. It has some structural repetition issues, and it loves its em-dashes and slop sentences way too much, but this is definitely going to be very strong base for fine-tuners. Also, its SWA KV cache means that you can fit a lot more context in much less space that you'd normally need.

Qwen/Qwen3.5-27B

Image-Text-to-Text • 28B • Updated Feb 25 • 3.23M • • 952

Note 2026's flavor of Qwen. Excellent for serious stuff, as usual. Plus at 28B, you can fit the vision layer without having to kill too much context or quantizing too hard. The thinking block tends to be massive, though.

zai-org/GLM-4.7-Flash

Text Generation • 31B • Updated Jan 29 • 683k • • 1.71k

Note Damn it's fast and clever. I'm generally not a fan of MoE models, but this one is really, really, good. It will however eat a bazillion tokens in its thinking block. It is very bad at writing, and quite censored.

zerofata/Q3.5-BlueStar-v2-27B

27B • Updated Mar 20 • 83 • 36

Note Creative model based on Qwen 3.5. Thinking and tool-calls are preserved (but model needs to be given orders). Strong contender for a mixed use case. Decent job, but still early. v3 is shaping to be a lot better.

zerofata/MS3.2-PaintedFantasy-v4.1-24B

24B • Updated Feb 19 • 37 • 11

Note First RP model of the year in this list. Similar to previous iterations, but grounded by normal assistant prompts. It's a really solid model, fun to use, adaptable beyond RP too. It uses L7-Tekken as an instruct format. Supports thinking mode optionally.