Reading path · 5 stops · 30 min of reading · 2 tools and data pages
Run models yourself.
What you'll learnWhen open weights pay off: browser inference, the RAM a giant model needs, and the parts of an inference stack nobody gives away.
The path, in order
Topics: LLMs, AI industry. Ticks show what you have read in this browser.
- Article · 14 minRun an LLM in the browser: WebGPU did 63 tokens/s, WASM did 1.6
How to run an LLM in the browser with Transformers.js, a Web Worker and WebGPU. Real code from a live app, three models measured, and the bundler trap.
- Free toolLLM VRAM calculator
Will the model fit your GPU? GB for weights + KV cache on 24 open models, and which cards fit.
- Article · 5 minMeta open-sourced an 816B model. Running it takes about 1TB of RAM
Meta released Muse Spark Open weights: 816B parameters, 42B active, 1M context. The model card asks 700GB VRAM at NVFP4, 2TB at BF16. What open means here.
- Article · 11 minDeepSeek gave away its engine. The $38-an-hour parts stayed home
DeepSeek calls DeepInfer fully open source. The README lists 3 missing parts, 4 of 5 routing boxes are absent, 6.82% matches vLLM, Huawei kernels 52% bigger.
- DataLLM API prices, and the cheapest model for your job
LLM API pricing for 71 models per 1M tokens, checked at the source. Pick for me: your job, volume and budget give 3 models with the monthly cost. Filters, CSV.
When you finish
Paths that pick up where this one stops.
AI API costs, and how to cut them
What models really cost per job, how to count tokens, and the caching, retry and proxy tricks that cut the bill.
AI news, checked against the sources
The AI stories that went around, with the filings, prices and code that confirm or sink them.