← Blog

Interactive explainers

Charts, calculators and simulations you can change, built on real numbers from papers, model files and production systems. Free for teaching under CC BY-SA 4.0.

22 figures

  1. Screenshot: Trimming touches two tables, and in a tied model they are the same one

    Diagram

    Trimming touches two tables, and in a tied model they are the same one

    Interactive diagram of Qwen3-0.6B: input embedding and tied output head drawn to scale against 28 unchanged layers, for the original vocabulary (151,669 rows), the language step (109,357) and frequency trims of 20% (87,498) and 60% (43,755).

    Vocabulary trimming, compared: what SqueezeBits measured on the phone, and the tokenizer check that makes it safe

  2. Screenshot: The bigger the vocabulary, the more trimming buys on a phone

    Interactive chart

    The bigger the vocabulary, the more trimming buys on a phone

    Gains in the decode-heavy setting (64 tokens in, 512 out) as stated in the SqueezeBits post.

    Vocabulary trimming, compared: what SqueezeBits measured on the phone, and the tokenizer check that makes it safe

  3. Screenshot: Drop one piece nobody “uses”, and the words built on it fall apart

    Simulation

    Drop one piece nobody “uses”, and the words built on it fall apart

    Interactive BPE merge tree for "there": without merge closure the parent "th" is dropped and the word splits into 4 tokens.

    Vocabulary trimming, compared: what SqueezeBits measured on the phone, and the tokenizer check that makes it safe

  4. Screenshot: Same compression, more broken text, and no warning before you ship

    Interactive chart

    Same compression, more broken text, and no warning before you ship

    Re-split: how much new text the rule splits differently, at the same number of rows removed, as a multiple of the merge-closed rule (median).

    Vocabulary trimming, compared: what SqueezeBits measured on the phone, and the tokenizer check that makes it safe

  5. Screenshot: Gemma 4 E2B, served in Thai and English: one fifth of the weights are rows you never select

    Annotated screenshot

    Gemma 4 E2B, served in Thai and English: one fifth of the weights are rows you never select

    The Where to cut panel after loading google/gemma-4-E2B-it and counting the model's own tokenizer.

    The vocabulary tax: the token tables your users never touch, and how to trim them safely

  6. Screenshot: The smaller the model, the bigger the share you never use

    Interactive chart

    The smaller the model, the bigger the share you never use

    Interactive bar chart: removable share of stored bytes for seven reference models, script rule against 15,000 Thai lines (paper Table 13).

    The vocabulary tax: the token tables your users never touch, and how to trim them safely

  7. Screenshot: A Thai user pays a memory tax or a compute tax. The tokenizer picks which.

    Interactive chart

    A Thai user pays a memory tax or a compute tax. The tokenizer picks which.

    Interactive scatter: rows unused by Thai and English against Thai-to-English token ratio for 19 tokenizer families (paper Table 1).

    The vocabulary tax: the token tables your users never touch, and how to trim them safely

  8. Screenshot: Each doubling of your text cuts the re-split rate by about a third

    Interactive chart

    Each doubling of your text cuts the re-split rate by about a third

    Interactive line chart: real re-split rate of new Thai text against trimming lines for four tokenizer families (paper Table 5).

    The vocabulary tax: the token tables your users never touch, and how to trim them safely

  9. Screenshot: Thai needs about 4,000 lines of real text. English web text needs 190,000.

    Interactive chart

    Thai needs about 4,000 lines of real text. English web text needs 190,000.

    Interactive dot plot: lines of own text needed for a 1% re-split rate in 20 languages (paper Table 9).

    The vocabulary tax: the token tables your users never touch, and how to trim them safely

  10. Screenshot: Look up any model before you download it

    Annotated screenshot

    Look up any model before you download it

    The vocabulary tax table on the Hub Weight Census page, Thai + English with 4,000 lines of text: Qwen/Qwen3-0.6B never uses 37.3% of the model, with 1.03% of words split differently.

    The vocabulary tax: the token tables your users never touch, and how to trim them safely

  11. Screenshot: Hide every name and the method still gets 95% of roles right

    Interactive chart

    Hide every name and the method still gets 95% of roles right

    Ground truth is the Python class of each module, which the method never sees.

    Reading a model's architecture from its safetensors header, and what 12,974 Hugging Face models reveal

  12. Screenshot: A 61 GB mixture-of-experts model, read from 2.3 MB of headers

    Annotated screenshot

    A 61 GB mixture-of-experts model, read from 2.3 MB of headers

    The Anatomy view after pasting Qwen/Qwen3-30B-A3B.

    Reading a model's architecture from its safetensors header, and what 12,974 Hugging Face models reveal

  13. Screenshot: Five parts, one stack, one loop per token

    Annotated screenshot

    Five parts, one stack, one loop per token

    The same model in the parts view, with the Weights lens on.

    Reading a model's architecture from its safetensors header, and what 12,974 Hugging Face models reveal

  14. Screenshot: In four years, MoE and grouped-query attention went from nothing to normal, and fewer new models fit a laptop

    Interactive chart

    In four years, MoE and grouped-query attention went from nothing to normal, and fewer new models fit a laptop

    Share of popular original language models created up to 2022 against those released in 2026.

    Reading a model's architecture from its safetensors header, and what 12,974 Hugging Face models reveal

  15. Screenshot: Tight clusters survive a shift that pushes loose ones over the line

    Simulation

    Tight clusters survive a shift that pushes loose ones over the line

    Interactive: synthetic emotion embeddings under a domain shift; tight angular clusters stay correct where softmax clusters cross boundaries.

    Why emotion models fall apart outside the lab (and what actually helps)

  16. Screenshot: Fifty concurrent voice sessions on an 8B model fit one 80 GB GPU

    Annotated screenshot

    Fifty concurrent voice sessions on an 8B model fit one 80 GB GPU

    The calculator with its default pipeline: Whisper large-v3, Llama 3 8B and XTTS-v2, 50 concurrent sessions, 4,096 tokens of context, FP16 weights.

    FinOps-Estimator: sizing GPUs for a voice AI pipeline

  17. Screenshot: The same 70B model needs 140 GB in bf16 and about 36 GB in 4-bit

    Interactive chart

    The same 70B model needs 140 GB in bf16 and about 36 GB in 4-bit

    N × bits ÷ 8, in GB of 10⁹ bytes.

    Fitting an LLM in your own server room: the memory math behind on-prem AI

  18. Screenshot: One large weight makes every other weight in its group less precise

    Simulation

    One large weight makes every other weight in its group less precise

    Interactive: 16 weights quantized to 4 bits with one scale; raising one outlier coarsens the grid for all others.

    Fitting an LLM in your own server room: the memory math behind on-prem AI

  19. Screenshot: Will it fit, and how fast can one user read it?

    Calculator

    Will it fit, and how fast can one user read it?

    Calculator: weights + KV cache + 12% runtime for Llama 3 8B/70B shapes against GPU memory, with the single-stream speed ceiling.

    Fitting an LLM in your own server room: the memory math behind on-prem AI

  20. Screenshot: A rotation turns one huge coordinate into sixteen ordinary ones

    Simulation

    A rotation turns one huge coordinate into sixteen ordinary ones

    Interactive: a 16-dimensional vector with one outlier before and after a Walsh-Hadamard rotation; the peak falls from 4.0x to 1.3x the RMS.

    TurboQuant, reviewed: when a theory paper walks straight into production

  21. Screenshot: About 800 ms from your last word to its first syllable, and end-of-turn detection is the biggest slice

    Calculator

    About 800 ms from your last word to its first syllable, and end-of-turn detection is the biggest slice

    Interactive latency budget: end-of-turn 250 ms, ASR 100 ms, LLM first token 300 ms, TTS 100 ms, network 50 ms against an 800 ms target.

    Your MAU is not your load: capacity math for voice AI at scale

  22. Screenshot: A million monthly users is about 150 LLM requests in flight at peak

    Calculator

    A million monthly users is about 150 LLM requests in flight at peak

    Calculator: monthly users to daily users, peak arrivals, live sessions (Little's law) and LLM requests in flight, then KV memory.

    Your MAU is not your load: capacity math for voice AI at scale