Tested Cogito V1 14B Qwen on my Linux server. 45 t/s, 9.7GB VRAM, and the same IDA self-awareness trick its 8B sibling pulled -- Run 2 deliberately stepped back to brute force because a beginner probably needed simpler first. Run 3 came back stronger with a nice candy analogy. That's DeepCogito's IDA training making a transformation of Qwen into something way better.
Read the full breakdown below.
Tested Cogito V1 8B on my Linux server. 83 t/s, 5.4GB VRAM, 131k context. The real story is where it deliberately wrote worse code because it decided a beginner needed simplicity over efficiency -- and admitted it! That's IDA self-reflection making a live call.
I guess a 5GB model with a conscience is worth more than a 70B model with none?
Read the full breakdown below.
Tested ServiceNow's Apriel 1.6 15B Thinker on my RTX 5060 Ti -- and the thinking logs made me put my tea down.
This model runs a compliance check before it writes a Python function. Literally. "We need to comply with the request. No disallowed content." Enterprise
DNA, fully intact.
But buried inside that corporate throat-clearing is something genuinely impressive. Full breakdown on the blog -- link in comments.
#AI #LocalLLM #Ollama #HomeLabAI #ReasoningModel #ServiceNow #OpenSourceA
Tested Mistral Small 3.1 24B on my RTX 5060 Ti -- 14.4GB VRAM, 20 tokens/sec, and barely 1.6GB of breathing room left on the card.
Slow? Yes. Worth it? Surprisingly, yes.
The output quality has no business being this good on a model pressed against the VRAM ceiling. Badly wish I had a better card. This thing's good!
Full benchmark breakdown on the blog -- link in comments.
Archer Dynamics
