> ## Content Index
> Fetch the complete content index at: https://claudefox.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Google patched glitch tokens in the Gemma 4 E2B and E4B tokenizers: how to check your copy
- URL: https://claudefox.com/gemma-4-e2b-e4b-tokenizer-glitch-token-fix/
- Published: 2026-11-16T21:31:00.000Z
- Updated: 2026-11-16T21:31:03.000Z
- Description: Google patched tokenizer.json for Gemma 4 E2B and E4B on November 3, 2026, making 217 to 233 glitch tokens unreachable so they stop garbling output. Weights are unchanged, and the 26B and 31B repos have not been patched yet.
- Author: hupicake
- Tags: Tools

If you run the small Gemma 4 models locally, the tokenizer you downloaded may be out of date. On November 3, 2026, Google merged tokenizer-only patches into four Hugging Face repositories, google/gemma-4-E2B, google/gemma-4-E2B-it, google/gemma-4-E4B and google/gemma-4-E4B-it. They remediate what the pull requests call glitch tokens: vocabulary entries whose embeddings have collapsed to near-identical representations, causing unpredictable model behavior when they appear in input.

The counts differ by model: 219 glitch tokens in E4B and E4B-it (253 BPE merges removed), 217 in E2B (246 merges removed) and 233 in E2B-it (267 merges removed). Most are fixed by deleting the merge rules that build them, so the same text is now split into well-trained subwords. Each E2B model also has one single-character glitch token, the rare character U+2127A (ID 249788), which is renamed to an unreachable placeholder; that character now encodes as UTF-8 byte tokens instead. In all four repos the vocabulary stays at 262,144 entries, token IDs and tokenizer\_config.json are unchanged, and no model weights were modified; the two instruction-tuned PRs also state the chat template is identical.

The pull requests show what the bad tokens did. Asked to repeat drawingCodehint, Gemma-4-E4B-it returned an empty string with the old tokenizer, which encoded it as the single token 226565\. Asked to repeat CYCLONEDB (token 192703), E2B answered with an unrelated Cyrillic word and E2B-it produced a thinking trace about a different string. With the patches, both strings split into two tokens and are repeated correctly. On E2B-it's repeat-the-glitch test across all 233 flagged tokens, the pass rate rose from 1.7% to 82.0%; the PR attributes the remaining failures to model capability limits on rare strings, not to the tokenizer.

A few ordinary vocabulary entries change too: 5 in E4B and E4B-it, 9 in E2B and 8 in E2B-it, whose only merge path ran through a removed glitch token, now decompose into subwords. The PRs report no behavioral regressions on those in their QA checks.

## What about Gemma 4 26B and 31B?

The PRs describe the work as part of a series of tokenizer fixes across all Gemma model repositories, Gemma 1 to 4 and MedGemma, 18 models in total, and warn that many remediated tokens are shared across Gemma 1, 2, 3 and 4\. On November 17, the google/gemma-4-26B-A4B-it and google/gemma-4-31B-it repositories had no such commit; their latest changes date from July 2026\. Not patched yet does not mean unaffected. The PRs recommend applying the patch consistently in agentic pipelines so these strings are not passed to unpatched models.

## We checked the E4B-it tokenizer

We downloaded tokenizer.json for google/gemma-4-E4B-it at the current main revision and at the previous commit and loaded both with the Hugging Face tokenizers library (version 0.23.3). The old file has 514,906 BPE merges and the new one 514,653, matching the pull request. drawingCodehint now encodes as \[140974, 45581\] instead of \[226565\]. AWSJavaScriptSDK, one of the downstream entries rather than a flagged glitch token, now encodes as \[55205, 44195, 36380\] instead of the single token 137313\. We did not load the model weights or test generation.

To see which version you have, run this in the folder that holds the tokenizer.json your tools actually load (for example your local model folder or Hugging Face cache snapshot):

```
python3 -c "import json; print(len(json.load(open('tokenizer.json'))['model']['merges']))"
# all four models before the fix: 514906
# patched: E4B and E4B-it 514653, E2B 514660, E2B-it 514639
```

## FAQ

Do I need new Gemma 4 weights? No. The patch only changes tokenizer.json; the pull requests state that no model weights are modified.

Does the patch change token IDs or vocabulary size? No. The vocabulary stays at 262,144 entries and token IDs are not re-indexed; the flagged entries simply become unreachable.

Are Gemma 4 26B and 31B patched too? Not as of November 17, 2026\. Their Hugging Face repositories show no glitch-token commit, although the PRs describe a series covering all Gemma models.

Source: [google/gemma-4-E4B-it pull request #50: Remediate 219 glitch tokens in tokenizer.json](https://huggingface.co/google/gemma-4-E4B-it/discussions/50?ref=claudefox.com), [google/gemma-4-E2B-it pull request #51: Remediate 233 glitch tokens in tokenizer.json](https://huggingface.co/google/gemma-4-E2B-it/discussions/51?ref=claudefox.com), [google/gemma-4-E2B pull request #13](https://huggingface.co/google/gemma-4-E2B/discussions/13?ref=claudefox.com), [google/gemma-4-E4B pull request #7](https://huggingface.co/google/gemma-4-E4B/discussions/7?ref=claudefox.com), [google/gemma-4-31B-it commit history](https://huggingface.co/google/gemma-4-31B-it/commits/main?ref=claudefox.com), [google/gemma-4-26B-A4B-it commit history](https://huggingface.co/google/gemma-4-26B-A4B-it/commits/main?ref=claudefox.com)

This story first appeared in the [claudefox daily briefing for November 17, 2026](https://claudefox.com/daily-2026-11-17/).