Author: Jiri Krecek
-
AI@Home – Phi4 14B Review
Conditions & Context Today we are investigating Microsoft’s serious foray into open LLM models, the Phi4 in 14B weight. An interesting iteration. I’m saying interesting, because it’s been known for a long while that Microsoft was closely working with OpenAI and the Microsoft Copilot is based on the latest versions of GPT. That did not…
-
Ministral 3 8B – Small Model. Big Personality. My favorite.
Model conditions & context Alright, so today we are taking a quick peek at Mistral’s latest open weight model, Ministral 3 in 8B size and with Q4 K M quantization. I was quite impressed with Mistrals 7B model and though the Ministral series is a different branch, I was still excited to give it a…
-
DeepSeek R1 8B – Lost in Time
Conditions & context This is a follow-up to my earlier AI@Home DeepSeek R1 8B article. If you haven’t read that one yet, go read it first — this one won’t make nearly as much sense without it. Are you back? Good. Because what happened next is genuinely fascinating and I did not plan any of…
-
Ministral 3 14B – Can Now See!
Model conditions & context Today we are looking at Mistral’s very fresh model (I think this is from December 2025) Ministral 3 14B in Q4 K M quantization variant in Instruct weight. So, this is a much younger model than the 12B Mistral I recently tested and really liked. Let’s see how it did and…
-
AI@Home – DeepSeek R1 8B Refactoring – This is the end!
Conditions & Context One of my readers on Mastodon asked an interesting follow up question to my DeepSeek R1 review: “Can it refactor?”. So, going back to this model, but this time we are not focusing on speed, but rather on refactoring: can a model take a sample of code, make it more robust, more…
-
AI@Home: Qwen 2.5 Coder 14B – Refactoring
Conditions & Context Today we are looking at Qwen2.5 Coder 14B and testing how good it is in refactoring some Python code. Many in the local AI community swear by this model and I wanted to see whether there was any truth to it or not. As a non-developer who is learning Python I have…
-
AI@Home – DeepSeek R1 8B
Conditions & context Today we are poking needles into DeepSeek R1 8B AI model, specifically the 0528 update, which is based on Qwen3 8B model. Since I have an 16GB card I’ve picked the 8-bit variant. Spoiler alert! This model would give Dostoyevsky and Tolstoy run for their money! :-) As in all my tests,…
-
AI@Home – Mistral Nemo 12B
Conditions & context Today we’re looking at Mistral Nemo 12B model with 5-bit quantization and if my great experience of their 4B model was any indication, this 3-times-larger model with 12 billion parameters is looking up! Let’s dive in! As in all my tests, I use the same prompt, the same hardware, and the same…
-
AI@Home – Gemma 3 12B
Conditions & context Today we are diving into a quick test of Google’s Gemma3, again a QAT quantized model, but this time a 12B. So, let’s hope that those 12 billion parameters do a better job than that “crazy” 4B sibling’s weights were. Let’s dive right in! As in all my tests, I use the…