Gradium's Phonon 31M is an English voice under Apache 2.0. The 40M models are not.
Gradium published Phonon 31M, an English text-to-speech model, under Apache 2.0. The 40M voices stay noncommercial and gated. The Rust runtime is a separate license.
The short version
Gradium's Phonon 31M is an English speech model under Apache 2.0, with the weights, tokenizer, and voices of that model in the grant. The 40M models are a different, noncommercial license. The Rust runtime is separate and is not a second-lab score.
What Gradium published
Gradium's project page is the record for Phonon 31M. It describes an English text-to-speech model, 31 million parameters excluding the audio tokenizer, mono audio at 24 kHz. The Apache 2.0 grant, on Gradium's account, covers the weights, the tokenizer, and the voices of that 31M English model. The default download the project points at is that English model. MarkTechPost's October 22 piece is a retelling of the same page. It is not a second measurement. Asked directly, Did a second lab score Phonon 31M? No score from a second lab is in these sources. MarkTechPost retells the project page. A benchmark this page did not run is not a result.what on-device inference keeps.
The size figure is the project's
Gradium says a q8 copy of the 31M weights is 64.5 MB. That number is the project's packaging note. This page did not weigh the file. A reader who needs the byte count should take it from the model card, and should treat a later upload as a later document. The parameter count and the file size answer different questions. One is the model. The other is a quantized file the project chose to publish.
The 40M models are a different grant
Gradium distinguishes the 40M models from the 31M English release. It says those models are under CC BY-NC 4.0 and that they are gated. A noncommercial grant is not Apache 2.0 with a different file name. Someone who downloads a gated 40M checkpoint is not holding the license this brief is describing. The wire item that says Phonon is Apache 2.0 is true of the 31M English model the project names, and it is not a blanket over every Phonon checkpoint.
The runtime is not the weights
Gradium says the runtime is written in Rust, that it does not need PyTorch, and that it streams audio while it generates. The project says that runtime is under MIT OR Apache-2.0, separate from the weight license. A runtime license does not relicense the voices. A weight license does not turn the player into a benchmark. The GitHub release the project points at for the runtime is a code drop. It is not a claim that a second lab reproduced a score.weights you can fetch and code you can read.
What this page will not paste
It will not paste command lines, checksums, or a sample transcript. Those belong on the project page. A news brief can name the model, the license, the language, and the split between the 31M grant and the 40M grant. It does not need to become an install guide. A reader who wants the file wants Gradium or the model card, not a rewritten flag list. Asked directly, Does the 31M count include the audio tokenizer? Gradium says the 31 million parameter count excludes the audio tokenizer. The figure is the project's, not a second lab's recount.
MarkTechPost is the other account
The October 22 MarkTechPost item is how the model reached the wire. It does not add a listening panel, a word-error table, or a license the project did not state. Where the retelling and the project page could be read two ways, this brief follows Gradium on the license split and on the parameter count. A later correction from Gradium would replace a sentence. A blog summary would not. Asked directly, Is the Rust runtime the same license as the weights? No. Gradium says the runtime is Rust, without PyTorch, under MIT OR Apache-2.0, and that it is separate from the weight license.Voxtral is a voice stack.
Who is in scope
People deciding whether the 31M English checkpoint is something they can ship under Apache 2.0. People who were handed a 40M file and told it was the same grant. The runtime is in scope only as a separate program with its own license. A phone that cannot run the file is a hardware question this page does not answer. No customer list is in the sources. Asked directly, What audio format does the project name? Gradium says the 31M English model is mono at 24 kHz. That is a format note, not a listening test.
What stays on the record
October 22, 2026 project page. Phonon 31M, English, mono 24 kHz, Apache 2.0 for that model's weights, tokenizer, and voices. Parameter count excludes the tokenizer. q8 size is Gradium's 64.5 MB figure. 40M models are CC BY-NC 4.0 and gated. Runtime is Rust, MIT OR Apache-2.0, separate. No second-lab score. MarkTechPost retells the page. The short answer this page will stand on is this: Gradium's Phonon 31M is an English speech model under Apache 2.0, with the weights, tokenizer, and voices of that model in the grant. The 40M models are a different, noncommercial license. The Rust runtime is separate and is not a second-lab score.what a license obligates.
What would move the row
A Gradium correction that changes the 31M license, or a second lab that publishes a score with its own method, would be a later document. A new voice pack under a new grant would be a new row, not a silent edit of this one. Until then the record is the project page and a retelling that does not add a measurement. Asked directly, Is Phonon 31M the same license as the 40M models? No. Gradium says the 31M English model is Apache 2.0. It says the 40M models are CC BY-NC 4.0 and gated.an evaluation a second lab can run.
Sources
The reports this brief is filing. Futureweb did not republish them.
Questions
Is Phonon 31M the same license as the 40M models?
No. Gradium says the 31M English model is Apache 2.0. It says the 40M models are CC BY-NC 4.0 and gated.
Does the 31M count include the audio tokenizer?
Gradium says the 31 million parameter count excludes the audio tokenizer. The figure is the project's, not a second lab's recount.
What audio format does the project name?
Gradium says the 31M English model is mono at 24 kHz. That is a format note, not a listening test.
Is the Rust runtime the same license as the weights?
No. Gradium says the runtime is Rust, without PyTorch, under MIT OR Apache-2.0, and that it is separate from the weight license.
Did a second lab score Phonon 31M?
No score from a second lab is in these sources. MarkTechPost retells the project page. A benchmark this page did not run is not a result.