Claude Code is becoming worse and worse

tokenmaxing is the stupidest thing I have ever heard

Gemma4-31B is workable, but two problems with GGUF version now: 1. prefill/intake speed too slow 2. if change the context and break KV cache may kill the llama.cpp service May need to upgrade the llama.cpp and rebuild it.

+1

After trying these and that, unsloth/Qwen3.5-27B-UD-Q4_K_XL.gguf is the best. Better than claude-opus 4.6 distilled, better than qwen3.5-122b-a10b, better than Gemma4-31b, better than Qwen3-Coder-Next

Microsoft is still put office in the center, AI as assistant. This is completly wrong. AI should be in the main stage, and Office as the tool. Anyway, it is just so hard for this giant to change the mindset, Open Mind! M$

Today OpenCode, tomorrow OpenClaw, another day Hermes ... endless. Folks, if you are really seriously using AI. Build you own one, and keep improve it, make it remember everything, can see, can hear, and can speak, and it is only yours. We can do it.

In the end of your PPT, don't say "Thank You"! Instead. "here is my email, x, linkedin, reach out to me" etc. "Thank You" indicates, you are SHOWing something, trying to convince . The other way continue the conversation, agree?

Proved

Wow

This will be good for Mac, DGX Spark, and AMD devices

The more gpu you buy, the more you save

The data is incorrect, using self build llama.cpp, run unsloth/Qwen3.6-35B-A3b-Q8_K_M on RTX 3090 (x2) can reach to 107t/s

Black Box and Intelligence - How a Comet Millions of Miles Away Changed What We Did on Earth — And What It Teaches Us About AI Some of my thoughts about Intelligence medium.com/p/a1bbe8bbb219

New best local model for y'all 16GB-64GB rejoice, the chosen one has arrived. huggingface.co/Qwen/Qwen3.6-2…

Unsloth is really fast huggingface.co/unsloth/Qwen3.…

Even time stops today, buy an RTX 3090 to run Qwen3.6-27B is still worthy

There has never been a better time to build a 2x RTX 3090s machine & get an LLM up and running locally