Blog

I tested 15 Qwen3.8 27B fine-tunes, quants and settings on two RTX 5070 Ti cards against the base model, plus a coding test built from my own projects. None beat it overall, but Tess-4.5 fixed as many of my bugs in about two-thirds of the time. Here's the two-model setup and settings I'd run on 32 GB of VRAM, for chat and for coding.

How Qwen3.8-27B works: a narrated, interactive 3D animation of Gated DeltaNet, attention and dense neural networks, using numbers recorded from the model.

Explore Z.ai's GLM-5.3 in an interactive 3D animation: how it chooses which parts of a conversation to read, builds an answer and guesses ahead to save work.

My Qwen3.8 27B benchmarks on a 16GB RTX 5070 Ti: base vs. Swift 1.5 for coding, Bonsai 2 for speed, with tested settings you can try.

Explore Xiaomi's MiMo-V2.6-Pro in an interactive 3D animation: how tokens, attention, a mixture of experts and speculative decoding produce an answer.

Take a night drive through Kimi K3 as an interactive 3D city: follow one token through Kimi Delta Attention, MLA, attention residuals and 896 experts.

Take the red pill and follow one token through DeepSeek V4-Pro in interactive 3D: compressed attention, hyper-connections, 384 experts and 61 layers.

Press start and watch a large language model finish a sentence in an interactive visualization: tokens, attention, the KV cache and mixture of experts.

How a machine that only predicts the next word ends up building world models, planning ahead, and doing things that look like thinking.

Six principles for using AI effectively: prepare the context, aim at goals instead of tasks, master the handoff, match the model to the task, let the AI triage your decisions, and tinker.