Ollama concurrency: NUM_PARALLEL and MAX_QUEUE
The second request waits, or it is refused: how OLLAMA_NUM_PARALLEL and OLLAMA_MAX_QUEUE decide which, and why every parallel slot costs you VRAM.
#vram
Filtering by topic #vram · clear
Ollama quantization: Q4 vs Q8 vs fp16
Pick an Ollama quantization with arithmetic instead of guesswork: what q4_K_M, q8_0 and fp16 cost in RAM, and where the quality actually drops.
What it takes to self-host Kimi K3
Kimi K3 is 2.8 trillion parameters. Here is the VRAM arithmetic, the KV cache math, and the three honest ways to run it without a 32 GPU cluster.
Self-hosted AI video generator: the reality
What open video models really produce in July 2026, the VRAM class each one needs, what a five second clip costs on a rented GPU, and when to use an API.