Bored, thought this screenshot was cute, might delete later.

Model: Kimi K2 thinking
Use case: idk it's just cool having a huge model running local. I guess I will use it for brainstorming stuff, medical stuff, other questionable activities like academic writing. PP speed/context size is too limited for a lot of agentic workflows but it's a modest step above other open source models for pure smarts
PP speed: Q3 GGUF 25 t/s (gguf 4 bit at 26k context) faster with lower context;
Token gen speed: 3ish to 20 t/s depending on context size

Model: GLM 4.6
Use Case: vibe coding (slow but actually can create working software semi-autonomously with Cline); creative writing; expository/professional writing; general quality-sensitive use
PP Speed: 4 bit MLX 50-70 t/s at large context sizes (greater than 40k)
Token Gen speed: generally 10-20

Model: Minimax-m2
Use case: Document review, finance, math,
PP Speed: MLX 4 bit 3-400 at modest sizes (10k ish)
Token gen speed: 40-50 at modest sizes

Model: GPT-OSS-120
Use case: Agentic searching, large document ingesting; general medium-quality, fast use
PP speed: 4 bit MLX near 1000 at modest context sizes. But context caching doesn't work, so has to reprocess every turn.
Token gen speed: about 80 at medium context sizes

Model: Hermes 405b
Use case: When you want stuff to have that early 2024 vibe... not really good at anything except maybe low context roleplay/creative writing. Not the trivia king people seem to think.
PP Speed: mlx 4 bit: Low... maybe 25 t/s?
Token gen Speed: Super low... 3-5 t/s

Model: Deepseek 3.1:
Use case: Used to be for roleplay, long context high quality slow work. Might be obsoleted by glm 4.6... not sure it can do anything better
PP Speed: Q3 GGUF: 50 t/s
Token gen speed: 3-20 depending on context size