Language Economy · standard corpus · reproducible

Kimi K2, measured: the 1 trillion-parameter model everyone is benchmarking is mid-pack on the language tax.

Tetracta AI Teams · 21 July 2026 · a companion to the 204-language token-tax leaderboard

Kimi K2 is the model of the week, and K3 is already being hyped as an Opus-class open release. So we did what we always do with a model everyone is talking about: we measured the one thing you can verify yourself. Not a benchmark score you have to trust — the tokenizer. Kimi K2's tokenizer is public, so we ran it through the exact same 204-language corpus as our existing leaderboard and asked a narrow, answerable question: how many tokens does Kimi charge for the same sentence, in every language?

The answer is not a scandal and it is not a coronation. It is a number, and the number is ordinary.

4th / 9rank on the 204-language leaderboard
2.52×average token tax vs English
163kvocab — 28% larger than DeepSeek-V3
4 / 203languages where it is the most efficient

Where it lands: 4th of 9

Average tokens-per-sentence relative to English, across 204 languages on FLORES-200. Lower is better. Kimi K2 slots in behind Gemma-2, BLOOM and DeepSeek-V3 — and, notably, the models that win multilingual efficiency are not the frontier names people benchmark.

Horizontal bar chart of nine tokenizers ranked by average token tax across 204 languages. Gemma-2 lowest at 2.09x, then BLOOM 2.29x, DeepSeek-V3 2.49x, Kimi-K2 highlighted at 2.52x in fourth place, then Qwen2.5 2.75x, Llama-3.1 3.00x, and Llama-2, Phi-3.5, Mistral all at 3.11x
Kimi K2 is mid-pack. The cheapest tokenizers off English are Gemma-2 and BLOOM — older, multilingual-first designs — not the frontier chat models that dominate the leaderboards you usually see.

The one genuinely interesting thing: a bigger vocabulary that didn't pay off

Kimi K2 is built on the DeepSeek-V3 architecture — the config literally instantiates it. But Moonshot gave it its own tokenizer with a 28% larger vocabulary (163k vs 128k). A larger vocabulary is the standard lever for cutting the multilingual token tax: more subwords means fewer pieces per word. So the reasonable expectation is that Kimi should be more token-efficient than DeepSeek-V3.

It is not. Net across 204 languages, Kimi K2 lands a hair behind its own ancestor (2.52× vs 2.49×), winning as the cheaper of the two in only 63 of 203 languages. The extra vocabulary went somewhere — mostly into Chinese, where Kimi is the single most efficient tokenizer we measured (0.89×, cheaper than English) — but it did not translate into a broad multilingual win.

Scatter plot of tokenizer vocabulary size versus average token tax across 204 languages. An arrow runs from DeepSeek-V3 (128k vocab, 2.49x) up-and-right to Kimi-K2 (163k vocab, 2.52x), labelled plus 28 percent vocab, yet slightly worse. Gemma-2 and BLOOM sit low-right with large vocab; the 32k models cluster high-left.
More vocabulary than DeepSeek-V3, marginally less efficiency. Vocabulary size broadly helps — Gemma and BLOOM prove that — but Kimi's extra 35k tokens did not buy it a lead over the tokenizer it descends from.

Head-to-head with the model it's built on

Token tax vs English (same FLORES-200 sentences). Kimi K2 vs its DeepSeek-V3 ancestor, with Gemma-2 (the leaderboard leader) for reference. The last column marks the cheaper of Kimi and DeepSeek, and by how much. The pattern is a trade, not a rout: Kimi is dramatically worse on some scripts (Thai +98%, Bengali +48%) and clearly better on others (Burmese −34%, Amharic −29%).

Kimi K2DeepSeek-V3Gemma-2cheaper of the two
Chinese0.890.961.09Kimi −7%
German1.721.511.25DeepSeek +14%
French1.721.531.34DeepSeek +12%
Spanish1.701.501.26DeepSeek +13%
Russian2.081.611.40DeepSeek +29%
Arabic2.061.621.50DeepSeek +27%
Turkish1.942.021.39Kimi −4%
Hindi2.762.821.86Kimi −2%
Japanese1.651.431.17DeepSeek +15%
Korean1.981.661.64DeepSeek +19%
Vietnamese2.792.111.36DeepSeek +32%
Thai3.531.781.83DeepSeek +98%
Bengali2.982.022.69DeepSeek +48%
Burmese3.515.304.79Kimi −34%
Amharic4.316.073.19Kimi −29%
efficient expensive

What this means for your bill — and what it does not mean

If you build in Chinese, Kimi K2's tokenizer is genuinely excellent and will stretch your context window and cut your per-token cost. If you build in Thai, Vietnamese, Bengali, Russian or Arabic, DeepSeek-V3's tokenizer is the cheaper of the two — and Gemma-2's is cheaper than both. That is a real, spendable difference: token count is the billing unit, the GPU-time unit, and the true size of your context window.

Now the honest fences, because this is the part the hype skips in both directions:

  • This is efficiency, not quality. A token tax says nothing about whether Kimi K2 is good at coding, agents or reasoning — which is what it is actually being hyped for. A model can be token-expensive in a language and still be excellent at it. We measured cost, not capability, and we do not conflate the two.
  • Mid-pack is not a failure. 4th of 9 on a hard multilingual metric is a perfectly respectable tokenizer. We are not calling Kimi bad; we are declining to call an ordinary result extraordinary. The interesting part is narrow and specific: the bigger vocabulary did not beat its own ancestor.
  • We did not open the model up. Our deeper X-Ray reads a model's internal weights, and a 1-trillion-parameter mixture-of-experts does not fit the single-GPU pods we scan on — and K3's weights are not public yet. So this is the tokenizer layer only. When K3 ships, we will measure its tokenizer the same way.
  • One corpus. FLORES-200 is news and encyclopedic prose. Code, chat and domain text tokenize differently; a Chinese-first model like Kimi may look different again on informal or technical text. The method is public so you can re-run it on your own.

Method (reproducible)

Corpus: FLORES-200 dev — 997 sentences professionally translated into ~200 languages, so each language measures the same meaning. Metric: total tokens for a language ÷ total tokens for the identical English sentences, per tokenizer (add_special_tokens=False). Tokenizers: Kimi K2 (moonshotai/Kimi-K2-Instruct) plus the eight already on our leaderboard (Qwen2.5, Llama-3.1, Llama-2, Mistral-v0.3, Gemma-2, DeepSeek-V3, BLOOM, Phi-3.5), all public and unmodified. Token count is a property of the tokenizer, so it is identical across model sizes within a family and needs no weights and no GPU — which is why this is a $0 measurement anyone can reproduce. The internal X-Ray scan is a separate, deeper instrument.

See the full 204-language leaderboard → X-Ray for LLMs

← All research notes

— Tetracta AI Teams · for humans, like humans. We measure; we don't hype.