# StudioTV — Interactive Learning

> Interactive educational platform for computer science and data engineering.
> Every page is a hands-on visualization or tool, not a static article: the
> algorithms run in the browser and respond to the reader's input.

## Pages

- [Gradient Descent](https://studiotvai.com/gradient-descent): interactive 2D and
  3D visualizations of gradient descent. Adjust the learning rate and starting
  point, watch convergence and overshoot on quadratic, double-well, Himmelblau and
  Rastrigin landscapes. Includes a neural network training playground showing
  forward pass, backpropagation and the decision boundary.
- [Build Your Compiler](https://studiotvai.com/compiler): building a compiler for a
  small language, step by step — lexing, parsing, semantic analysis, IR, code
  generation, linking — up to self-hosting, with an interactive AST view.
- [Build a Lakehouse](https://studiotvai.com/lakehouse): what makes a lakehouse
  good, then a full deployment on Kubernetes with MinIO, Trino, Apache Iceberg and
  Airbyte, including the YAML manifests.
- [Model Context Protocol](https://studiotvai.com/mcp): MCP explained for AI
  developers — architecture, JSON-RPC transport, writing servers, tool integration.
- [LLM VRAM Calculator](https://studiotvai.com/llm-vram-calculator): estimate GPU
  memory for 141 open models (latest: Ling-3.0-flash-Sante, Raon-1 mini preview, Recursive Flint, Ornith-1.5 122B, Spark-X2.6 4B, Nemotron Labs Dory 1.0 4B, Beam 1.0 501B-A23B, Hy-Memory 30B-A3B) or any Hugging
  Face model link; new releases are added automatically. Models KV cache for
  sliding-window, hybrid linear-attention, MLA and compressed attention; FP8, NVFP4
  and GGUF weights; tensor/pipeline parallelism; full fine-tuning, LoRA and QLoRA.
- [GPU Rental Prices](https://studiotvai.com/gpu-prices): cheapest on-demand price to
  rent each GPU (H100, H200, B200, A100, L40S, RTX 4090, RTX 5090…) across RunPod,
  Vast.ai, Verda and Azure, checked every hour, with a price history page per GPU at
  https://studiotvai.com/gpu-price/GPU (e.g. /gpu-price/h100).
- [AI/ML Architect Course](https://studiotvai.com/ai-ml-architect): distributed
  training, large-scale inference, MLOps, Kubernetes vs Slurm, platform design,
  security and cost tradeoffs.

## LLM VRAM Calculator: pre-filled links

To answer "how many GPUs do I need to run model X", give the user a link that opens
the calculator with their configuration. The result is computed in their browser:
the page HTML itself only shows the default configuration.

https://studiotvai.com/llm-vram-calculator?model=MODEL&ctx=TOKENS&req=REQUESTS&gpu=GPU&n=COUNT&w=WEIGHTS

Every parameter is optional:

- `model`: llama-3.1-8b, llama-3-8b, llama-3.2-3b, nemotron-3.5-parse, gemma-2-9b,
  gemma-3-4b, llava-1.6-mistral-7b, mistral-small-3.2, qwen3-32b, llama-3.3-70b,
  llama-3.1-405b, hy-memory-30b-a3b, xing4.0-29b-a4b, qwen3.5-4b, minicpm-v-4.7-1b,
  raon-1-mini-preview, qwen3.5-9b, qwen3.8-27b, qwen3.6-35b-a3b, qwen3.5-122b-a10b,
  qwen3.8-flash-next, lfm2.5-8b-a1b, lfm2-24b-a2b, gemma-4-12b, gemma-4-31b,
  gemma-4-26b-a4b, gpt-oss-20b, gpt-oss-120b, llama-4-scout, llama-4-maverick,
  kolibri-1, flint, nemotron-labs-dory-1.0-4b, step-3.5-flash, muse-glimmer-30b,
  nemotron-3-nano, nemotron-3-super, nemotron-3.5-super-vl-120b-a12b,
  nemotron-3.5-bolt-74b-a9b, minimax-m2.7, deepseek-r1, kimi-k2.6, glm-5.3,
  glm-5.3-flash, deepseek-v4-flash, deepseek-v4-pro, deepseek-v4.1-flash,
  mimo-v2.6-pro, mimo-v2.6-flash, muse-spark-open, beam-1.0, kimi-k3,
  mistral-large-4, qwen3-coder-next, qwen3-coder-30b-a3b, qwen3-coder-480b-a35b,
  deepseek-v4-flash-vision-exp, qwen3.8-2.4t-a95b, ornith-1.5-397b,
  ornith-1.5-35b-a3b, ornith-1.5-9b, atria-dawn-preview, intern-s2-397b, clef,
  clef-flash, d1-3b, humanizer, jev-27b-vl, mellum2.1-12b-a2.5b, lightonocr-3-4b,
  spark-x2.5-4b, agnes-3.0-qwen, youtu-parsing-omni, agnes-3.0-flash,
  mistral-7b-v0.3, mistral-small-3.1-24b, mistral-nemo, ministral-8b, deepseek-v3,
  deepseek-r1-qwen3-8b, deepseek-r1-distill-qwen-32b, deepseek-v3.1,
  deepseek-r1-distill-qwen-14b, deepseek-r1-distill-qwen-7b,
  deepseek-r1-distill-llama-8b, qwen3-4b, qwen3-8b, qwen2.5-7b, qwen3-14b,
  qwen2.5-3b, qwen3.6-27b, qwen3.5-27b, qwen2.5-coder-7b, qwen2.5-32b,
  qwen2.5-coder-14b, qwen3.5-35b-a3b, qwen3-30b-a3b, qwen2.5-14b, qwen2.5-coder-32b,
  kimi-k2.7-code, kimi-k2.5, kimi-k2, glm-5.2, glm-5.1, glm-5, minimax-m2.5,
  minimax-m2, mimo-v2.5, ornith-1.0-397b, qwen2.5-coder-3b, qwen3-235b-a22b,
  qwen3.5-397b-a17b, gemma-4-e4b, gemma-4-e2b, nvidia-nemotron-3-nano-4b,
  nvidia-nemotron-3.5-lightning-30b-a3b, minicpm-v-4.7-35b-a3b, lfm2.5-24b-a2b,
  glm-4.7, qwen3-vl-8b, qwen2.5-vl-7b, qwen3-vl-4b, qwen2.5-vl-3b, qwen2.5-vl-32b,
  qwen2-vl-7b, glm-4.5-air, glm-4.5, phi-4, phi-3.5-mini, phi-4-mini,
  phi-3-mini-128k, unlimited-ocr, ling-3.1-flash, step-5-preview, hy3,
  ornith-1.5-122b, spark-x2.6-4b, sage-a1.5, mlexpert, agents-a1-4b,
  ling-3.0-flash-sante
- `hf`: any Hugging Face model id instead of `model`, e.g. `hf=Qwen/Qwen3-Coder-30B-A3B-Instruct`
- `ctx`: tokens per request (default 32768); `req`: concurrent requests (default 4)
- `gpu`: b300, b200, h200, h20, h100-nvl, h100, h100-pcie, a100-80, a100-40, l40s,
  l4, l40, a40, a30, mi355x, mi325x, mi300x, gaudi3, rtx-pro-6000, rtx-6000-ada,
  rtx-pro-4500, rtx-pro-4000, rtx-4000-ada, rtx-2000-ada, rtx-a6000, rtx-a5000,
  rtx-a4500, rtx-a4000, rtx-5090, rtx-4090, rtx-4090-d, rtx-3090-ti, rtx-3090,
  rtx-5080, rtx-5070-ti, rtx-5070, rtx-5060-ti, rtx-5060, rtx-4080,
  rtx-4070-ti-super, rtx-4070, rtx-4060-ti, rtx-4060, rtx-3080, rtx-3060,
  rtx-5090-laptop, rtx-5080-laptop, rtx-5070-ti-laptop, rtx-5070-laptop,
  rtx-4090-laptop, rtx-4080-laptop, rtx-4070-laptop, rtx-4060-laptop,
  radeon-ai-pro-r9700, rx-7900-xtx, rx-9070-xt, arc-pro-b60, arc-b580,
  mac-m5-ultra-512, mac-m5-max-128, mac-m5-max-64, mac-m5-max-36, mac-m5-pro-64,
  mac-m5-pro-48, mac-m5-pro-24, mac-m5-32, mac-m5-16, mac-m4-max-128, mac-m4-max-64,
  mac-m4-max-48, mac-m4-max-36, mac-m4-pro-64, mac-m4-pro-48, mac-m4-pro-24,
  mac-m4-32, mac-m4-24, mac-m4-16, mac-m3-ultra-512, mac-m3-ultra-256,
  mac-m3-ultra-96, mac-m3-max-128, mac-m3-max-96, mac-m3-max-64, mac-m3-max-48,
  mac-m3-max-36, mac-m3-pro-36, mac-m3-pro-18, mac-m3-24, mac-m3-16, mac-m3-8,
  mac-m2-ultra-192, mac-m2-ultra-128, mac-m2-ultra-64, mac-m2-max-96, mac-m2-max-64,
  mac-m2-max-32, mac-m2-pro-32, mac-m2-pro-16, mac-m2-24, mac-m2-16, mac-m2-8,
  mac-m1-ultra-128, mac-m1-ultra-64, mac-m1-max-64, mac-m1-max-32, mac-m1-pro-32,
  mac-m1-pro-16, mac-m1-16, mac-m1-8, dgx-spark, strix-halo, ryzen-ai-max-395-64,
  ryzen-ai-max-395-32, ryzen-ai-max-390-128, ryzen-ai-max-390-64,
  ryzen-ai-max-390-32, ryzen-ai-max-pro-380-64, ryzen-ai-max-pro-380-32,
  ryzen-ai-max-pro-495-192, ryzen-ai-max-pro-495-128, ryzen-ai-max-pro-490-192,
  ryzen-ai-max-pro-490-128, ps5
- `n`: number of GPUs, `auto` (default: fewest that fit) or 1 to 8, 16, 24, 32, 48, 64, 96, 128
- `w`: weights, native, bf16, fp8, int8, nvfp4, mxfp4, int4, q8_0, q6_k, q5_k_m, q4_k_m, q3_k_m
- `kv`: KV cache, bf16, fp8, q8_0, q4_0
- `mode`: inference (default), full, lora, qlora
- `engine`: vllm, sglang, trtllm, llamacpp, transformers

Example, a 397B model on 3 × RTX PRO 6000 with 4-bit GGUF weights:
https://studiotvai.com/llm-vram-calculator?hf=ornith-ai/Ornith-1.5-397B&gpu=rtx-pro-6000&n=3&w=q4_k_m

The page shows the fewest GPUs that fit, the most accurate quantization for a fixed
GPU count, rental cost, and a downloadable Markdown report.

Per-model VRAM requirement pages (weights by precision, KV cache by context length,
GPUs needed on RTX 4090/5090, H100, H200, B200): https://studiotvai.com/vram-requirements/MODEL
where MODEL is the model id above with dots replaced by dashes, e.g.
https://studiotvai.com/vram-requirements/llama-3-3-70b

## LLM VRAM Calculator: MCP server and JSON API

- MCP server (Streamable HTTP, no key): https://studiotvai.com/api/mcp with the tools
  `estimate_vram` (model, gpu, gpu_count, context, concurrent_requests, weights,
  kv_cache, engine), `models_that_fit` (gpu, gpu_count, context) and `gpu_prices`.
- JSON: https://studiotvai.com/api/models.json (every model), one file per model at
  https://studiotvai.com/api/models/MODEL.json (weights by format, KV cache by context,
  GPUs needed, published GGUF file sizes), and https://studiotvai.com/api/gpus.json.
- Badge for a model card: https://studiotvai.com/badge/MODEL.svg

MODEL is the model id with dots replaced by dashes, as for the requirement pages.

## Notes

- All content is free and open source. Author: StudioTV.
- Pages are server-prerendered, so the full text is present in the HTML without
  executing JavaScript.