Tag

#llm

Every story tagged llm, newest first.

A GPU graphics card circuit board lying on a wooden workbench with tools and electronics visible in the blurred background

Which local LLMs fit in 8, 12, 16, 24 or 32GB of VRAM

At Q4_K_M and 8K context, 8GB comfortably holds Gemma 2 9B (5.8 GB) and 24GB holds Gemma 3 27B (17.5 GB). Full tables for every common VRAM tier, computed the same way our GPU checker computes them.

Ahmad J · Sep 6, 2026 · 8 min read

MacBook laptop on wooden desk by window with coffee mug, glasses, notebooks, and potted plant in home office setting

Tutorial · software

Evaluate whether a model is good enough for your task

Stop guessing from vibes. A repeatable way to decide if a model clears the bar for your specific job, using your own data.

Ahmad J · Sep 4, 2026 · 3 min read

A small square microchip sits on a concrete ledge in a blurred modern interior with glass partitions and red lighting

Guide · aiDeep read

Choosing the right model size for your task

Bigger is not automatically better. A decision framework for matching model size to the job, the latency budget, and the hardware you actually have.

Ahmad J · Sep 4, 2026 · 4 min read

Metal letterpress type blocks arranged in a wooden compartmented tray, with some blocks spelling partial text in sharp focus

News · ai

Claude Fable 5.1 costs 25% less, and its per-token price did not move

Anthropic's new model is cheaper for the reason nobody put in the headline: input and output rates are unchanged, and the whole discount is a 75 per cent cut to cache reads. That makes the saving conditional on how much context your workload re-reads, and the published benchmark gains are just as uneven.

Ahmad J · Sep 2, 2026 · 5 min read

Computer workbench with rows of RAM modules and an SSD displayed on a gray surface, with electronics and cables in the background

News · ai

DeepSeek's first V4 vision model is MIT-licensed, and 307 GB

DeepSeek shipped an open-weight multimodal model under MIT this weekend. Its own files carry three things the coverage left out: the download size, the fact that every benchmark is self-reported, and a footnote saying two of the wins are over a model that cannot see.

Ahmad J · Sep 1, 2026 · 5 min read

Which local LLMs fit in 8, 12, 16, 24 or 32GB of VRAM

Evaluate whether a model is good enough for your task

Choosing the right model size for your task

Claude Fable 5.1 costs 25% less, and its per-token price did not move

DeepSeek's first V4 vision model is MIT-licensed, and 307 GB

More stories