One file of commented source code costs qwen3-coder-30b 210 tokens in English and 353 in German. That German figure is the highest in our catalogue, 20% above the cheapest model measured, and it needs 17.28 GiB in Q4_K_M before any context is reserved.
It is not alone at the bottom, and that is the interesting part: three other models produce exactly 353, because all four share one tokenizer.
The same code, measured on 18 tokenizers
English is the baseline column, and the other three are measured against it. German is the column a Berlin team pays.
| Model | Size (reference quant) | English | German | vs the cheapest |
|---|---|---|---|---|
llama4-scout-17b |
60.87 GiB | 211 | 295 | — |
gpt-oss-20b |
10.83 GiB | 210 | 298 | +1% |
qwen3.8-27b |
15.33 GiB | 223 | 325 | +10% |
mistral-small-24b |
13.35 GiB | 217 | 327 | +11% |
gemma4-12b |
6.63 GiB | 239 | 332 | +13% |
glm-5.2 |
433.83 GiB | 209 | 343 | +16% |
deepseek-v4-flash |
127.28 GiB | 219 | 351 | +19% |
qwen3-8b |
4.68 GiB | 210 | 353 | +20% |
qwen3-coder-30b |
17.28 GiB | 210 | 353 | +20% |
Measured by Local AI Scope. Parallel corpus sha256[:16] 2f6c96b6ea0b161f, tokenized 2026-08-25, add_special_tokens=False. Size is the published Q4_K_M artefact, or the nearest compression the author publishes (deepseek-v4-flash at IQ4_XS).
Four models return 353: qwen3-1.7b, qwen3-4b-2507, qwen3-8b and qwen3-coder-30b, all on the same 151,936-token vocabulary. qwen3.8-27b, same vendor, newer tokenizer at 248,320 tokens, lands 28 tokens lower on the same file.
Of those four, qwen3-coder-30b is the only one built specifically for programming, and by far the largest: 17.28 GiB against 4.68 for qwen3-8b. A team choosing it is paying for the coding specialisation and inheriting the old tokenizer with it.
In English the difference nearly vanishes
The English column spans 209 to 239 tokens, a 14% spread, and qwen3-coder-30b sits near the efficient end of it at 210. Rank these models on English code alone and it looks excellent.
That is the transplant problem. The spread that matters to a German team lives in the German column, and it is wider than the English one: 295 to 353 is 20%, against 14% in English.
What it costs in French and Spanish
| Language | qwen3-coder-30b |
Cheapest measured | Gap |
|---|---|---|---|
| English | 210 | 209 | +0% |
| Spanish | 269 | 252 (gpt-oss-20b) |
+7% |
| French | 309 | 277 (gpt-oss-20b) |
+12% |
| German | 353 | 295 (llama4-scout-17b) |
+20% |
Measured by Local AI Scope, same corpus and same file.
A French team pays a 12% surcharge on this model. A German team pays 20%, on a model that also asks for 17.28 GiB when gpt-oss-20b sits one percentage point off the best score at 10.83 GiB.
What this does not say
Fewer tokens does not mean better code. We have measured what the same work costs on each tokenizer; we have not run these models or evaluated their output, because that needs a GPU and an evaluation protocol we do not have. gpt-oss-20b is not “a better coding model” on this evidence. It is a cheaper one per unit of German source code, which is a different and narrower claim.
Calling a formula a benchmark is the error this site exists to avoid, so the claim stops where the measurement stops.
The consequence for the machine, not just the invoice
Tokens are also how context memory is allocated, so the surcharge lands twice. A repository excerpt that fits an 8k window in English needs roughly 13.8k in German on this tokenizer, and that extra context competes with the weights for the same memory.
qwen3-coder-30b in Q4_K_M is 17.28 GiB before any context, which rules it out on every 8 GB and 12 GB machine in our profiles and makes it a 24 GB proposition. The effect shows up on small documents, not just long ones: on a PC with an RTX 4060 and 4,000 words of code at public sensitivity, English, Spanish and French all resolve to running locally, while German resolves to the cloud. Same machine, same task, same file.
What to check before accepting a coding recommendation
Look up the vocabulary size of the tokenizer rather than the release date or the family name. In this catalogue, models with vocabularies above 200,000 cluster between 1.29x and 1.33x for German when the four document types are taken together; those below 155,000 pay between 1.40x and 1.62x. Broken out by document type the two bands overlap, so this is a shortlist heuristic, not a per-document rule. qwen3-coder-30b sits at the far end of the second group despite being a 2026 release.
Then check the fit for your actual machine and language rather than the headline parameter count, since the context has to fit next to the weights. The Model Finder does both, the token cost table has the per-language ratios for all 18 models, and the runtime profiles cover the program that will load it.