Demietrich Baker

Software dev, local LLM lover

[Demietrich Baker @demietrich]

The final shootout - shoutout @nanbeige for birthing the current best small LLM.
Nanbeige4.2-3B is our winner 👑

Mad respect to our runners up @SparkLLM and @TheInclusionAI . Both amazing models, and honestly more truly "small"

https://t.co/ezDWLXR2jQ

[Demietrich Baker @demietrich]

Came home to my laptop having been pulled off the counter by my vacuum. Hooray

[Demietrich Baker @demietrich]

@notkingofchids I was surprised how big the gap was too. Plenty of it's failures were verbosity. I had theorized it to be the serving config, but we'd see the same gap with 4B if that were the case. Likely worth a re-test now that Spark X2.5 support is in llama.cpp main (gguf)

[Demietrich Baker @demietrich]

Small, accessible LLMs are crucial for local use. Which is most worth using?

Tested:
- Spark-X2.5-1.7B
- MiniCPM5-2B
- LFM2.5-2.6B
- Nanbeige4.2-3B
- Spark-X2.5-4B
- Ling-3.0-tiny
- LFM2.5-8B-A1B
- "Qwen3.8-9B-Distill"

https://t.co/VNm3ENC0rH

[Demietrich Baker @demietrich]

boy LLMs are neat huh?