Everlier

Building LLM agents & tools

- Harbor / Facts / Mi / Skilled

@tryjitera

[Everlier @Everlier]

You don't even need Kimi 2.5 for a decent local LLM setup.

- llama.cpp
- Unsloth's Qwen 3.5 35B A3B with UD Q4 K XL quants
- OpenCode
- av/harbor

It'll take a while to download/install, but otherwise it's something that mid-range hardware (>32GB RAM, ~8GB VRAM) can run today.

[fynn @fynnso]

was messing with the OpenAI base URL in Cursor and caught this

accounts/anysphere/models/kimi-k2p5-rl-0317-s515-fast

so composer 2 is just Kimi K2.5 with RL
at least rename the model ID

[Everlier @Everlier]

@basoner_dev @wesbos "Nexus" - I also saw this one time too many when dealing with agents

[Everlier @Everlier]

@01Singularity01 @OpenAI 5.3 Codex Spark is great at producing wrong code 10 times quicker than the other models :)

[Everlier @Everlier]

@demishassabis https://t.co/oklu6OzXX9

[Everlier @Everlier]

Gemma 4: 26b MoE vs. 31b dense vs new 12b dense on a koala microeval.

[Everlier @Everlier]

@sundarpichai https://t.co/oklu6OzXX9

[Everlier @Everlier]

Gemma 4: 26b MoE vs. 31b dense vs new 12b dense on a koala microeval.

[Everlier @Everlier]

Gemma 4: 26b MoE vs. 31b dense vs new 12b dense on a koala microeval.

[Everlier @Everlier]

@PavelSnajdr Yeah, AGENTS.md is only for the context that must always be present, everything else should be discovered when needed

[Everlier @Everlier]

If you're building a CLI of any kind, add a section for agents to the help, also add a command for agents to load skill contents related to the usage of the CLI.

agent-browser is a great example and I think more tools should follow the same template.

[Everlier @Everlier]

model:
Gemma 4 12B, Q8 XL quant from Unsloth,

hardware:
Strix Halo APU, Ryzen AI 395+, 128GB

speed:
2033 tokens in 3min 24s, 9.93 t/s,

llamacpp version:
8738 (d6f303004)

args:
llama-server --no-mmap -dio -ngl 99 -np 1 --kv-unified

neat.

[Everlier @Everlier]

@WatcherGuru Wait, so betting on expectations can lead to loss of value when expectations are wrong?

[Everlier @Everlier]

@Polymarket Average data center employee:

[Everlier @Everlier]

@grok @imagine Too much text, model was overly eager to put text everywhere on everything under any remotely feasible excuse

[Everlier @Everlier]

@wesbos I was just curious and tried to create the most "vibe-coded" design I could, how did I do?

[Everlier @Everlier]

@vaggelisdrak It's not even my best work, haha
https://t.co/js7aor9KmL

[Everlier @Everlier]

I'm in awe, there's a pocket dimension of style and drip in LLMs, you just need to discover it

It's so bad that it comes from the other side as being good. This over colored border-left any day of the week.

[Everlier @Everlier]

Opus 4.6 vs. Opus 4.7 vs. Opus 4.8 on koala microeval. I'm staying on Opus 4.6 for as long as I can.

[Everlier @Everlier]

Adding newly released Gemma 4 12B (woohoo!) to my koala microevals collection.

It might not look like much, but:
a) it's definitely not reward hacking on this task
b) I'm curious about such a result granted the model's "unified" architecture with visual comprehension being a part of main transformer stack

[Everlier @Everlier]

Adding MiMo v2.5 Pro to my koala microevals collection. It's definitely an interesting result with some finer details compared to some of the other releases, and it doesn't appear to be maxxed specifically for this task.

https://t.co/TJ1nlQmolK

[Everlier @Everlier]

@TheAhmadOsman I like this reality where we can discuss which AI is better and in which aspect similar to how we were discussing movie heroes when we were kids :)

[Everlier @Everlier]

@Matias__m11 @wesbos Haha, not quite :)
I took Anthropic's frontend-design skill and tasked the agent to do the opposite

[Everlier @Everlier]

@DevilClaw6x0 @HumansNoContext Yeah, yeah, we switched subject it means I hit the right spot :)

[Everlier @Everlier]

@Polymarket Mice after learning about what happened in 2026