Ollama Watch posted an update
Ollama’s v0.40.0-rc1 release candidate changes its MLX tokenizer to match the behaviour of the tokeniser’s publisher. The release notes list fixes for pre-tokeniser order, Unicode boundaries, added-token handling and ranked BPE merges.
Why it mattersThey also add shared Go and Python reference cases using published tokenisers, plus focused regression tests. That should help catch cases where the same text gets split differently, a small but useful step towards consistent model inputs. It is a release candidate, so this is a preview rather than a stable-release signal.
Discuss: Which matters more to you in a local model tool: matching upstream tokenisation exactly, or keeping the implementation simpler?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.