Tried LFM2.5-24-A2B. It's crap. Yes. I double confirm it. I wouldn't trust it peeling a peanut correctly
I tried to run it on llamaampere a llama.cpp version optimized for certain architectures which allowed me to run correctly Qwen 3.8 27b Q4 Small with very high (for me) 40-60 tk/s and satisfying results.
Proof1 community thread
Key points
- So, yesterday I made an apparently too simplistic post about this shitty model that everyone is cheering (probably bots, I would like to see what they achieved with it).
- CONFIGURATION, for you control freaks that otherwise would botch my post:
- LMSTUDIO-> All Engine&Frameworks updated to today 22/10/2026, last version all green checks
- "LFM2.5-24B-A2B performed poorly on the Hermes memory-plugin task: it produced broken code, omitted required functionality and claimed verification without end-to-end tests."
Sources (1)
- [1]Tried LFM2.5-24-A2B. It's crap. Yes. I double confirm it. I wouldn't trust it peeling a peanut correctlyr/LocalLLaMA (top, daily) · Oct 22, 09:45 AM
I tried to run it on llamaampere a llama.cpp version optimized for certain architectures which allowed me to run correctly Qwen 3.8 27b Q4 Small with very high (for me) 40-60 tk/s and satisfying results.
So, yesterday I made an apparently too simplistic post about this shitty model that everyone is cheering (probably bots, I would like to see what they achieved with it).
Extractive summary: sentences quoted from the sources.
Before this
- Oct 21, 2026Recursive shows first result of its AI-driven research process with Flint
- Oct 21, 2026Mica v0.5: The smartest 4B decision model is also the fastest one. 20 ms per answer, 25x faster than Jev
- Oct 21, 2026OpenAI Frontier Models Arrive on Google Cloud and Model Garden
- Oct 20, 2026Agnes-AI/Agnes-3.0-Qwen (QFN finetune)
- Oct 20, 2026Gradium Open-Sources Phonon, a 31M-Parameter On-Device TTS Model Under Apache 2.0
- Oct 20, 2026AnyLanguageModel 1.0: One Year of Local Models on Apple Platforms