Microsoft: Phi 4 is Microsoft's model. Its measured tokenizer sits with the OpenAI cl100k group, which is evidence that it builds on the OpenAI cl100k tokenizer. Tokenizer reuse is normal engineering, open vocabularies travel between labs, and this says nothing against Microsoft's authorship of the model itself.
cl100k_base is an open vocabulary table anyone can build on. Sharing it says which dictionary the model reads with, and nothing about who made the model.
Phi-4's repository tokenizer, run over the 50 strings offline, matches OpenAI's open cl100k_base encoding on every string (vocabulary 100,352, the padded cl100k table). IBM's Granite 4.0 and 4.1 declare the same vocabulary size and their endpoints return identical counts, which is why Microsoft and IBM share an exact signature group: both build on the tokenizer OpenAI released as open source with tiktoken.
Phi-4 repository (config.json: vocab_size 100352)Granite 4.0 H Micro repository (config.json: vocab_size 100352)OpenAI tiktoken (cl100k_base, open source)
| Signal | Result | Reference |
|---|---|---|
| Tokenizer signature | Exact match with 2 other models | tk_ba918336 |
| API surface | No other entry shares this exact contract | api_7f4ca365 |
| Context and output | 16,384 context · 16,384 max output | declared |
| Reasoning contract | None declared | declared |
| Serving providers observed | DeepInfra | measured 21 August 2026 |
Click a row to open the full comparison.
An x marks a row excluded as corrupt (caching or provider interference detected during measurement). Overhead subtracted: 7 prompt tokens.
[Microsoft Research](/microsoft) Phi-4 is designed to perform well in complex reasoning tasks and can operate efficiently in situations with limited memory or where quick responses are needed. At 14 billion...
Supported parameters
frequency_penaltylogit_biasmax_tokensmin_ppresence_penaltyrepetition_penaltyresponse_formatseedstopstructured_outputstemperaturetop_ktop_p
defaults: {}
Cite this page as the evidence record for Microsoft: Phi 4: the URL is stable, measurements are dated, and revisions append to the history above. Method: tokenizer fingerprinting, v1, 50 strings. The raw data behind every figure is on the data page.
