The PixelRaise blog
Sovereign AI, open source language models, the real performance of local and production AI, compliance: our analyses, with sources.
Articles
Mistral Large 4: should you self-host it?
Around 1 trillion parameters, with weights promised for late October. How many GPUs it takes, and when to host it yourself rather than use the API.
Read the articleSovereign RAG: the model matters less than your documents
Chunking, embeddings, re-ranking, access rights, evaluation: the 7 links where the quality of an internal AI is really decided, long before the choice of LLM.
Read the articleWhat does an internal generative AI really cost?
The 5 real cost lines, the break-even point against per-seat licences, and the 3 calculation mistakes that distort every comparison.
Read the article6 alternatives to Ollama for running an LLM locally
LM Studio, llama.cpp, KoboldCpp, Jan, vLLM, Msty: which one to choose for which use, and when to leave the workstation for the sovereign server.
Read the articleLocal GenAI on Apple Silicon Macs: what real performance?
MLX-LM plus OpenCode: architecture, throughput in tokens/s, unified memory, and how far local goes against a sovereign server deployment.
Read the articleMistral Small 4 vs Small 3.2 24B: what the new generation changes
A comparison with numbers: coding benchmarks, vision, inference throughput. Should you migrate to Small 4 for a sovereign AI in production?
Read the articleDemonstration
A sovereign AI for your organisation?
Let us move from theory to demonstration, on your own documents.