Breeze TTS 2 tops the open-weight TTS board under a licence that bars commercial use

Two artefacts, two licences. The inference code on GitHub is Apache 2.0, so it can be used, modified and redistributed commercially ^[5]. The weights on Hugging Face fall under the BreezeBlue Research and Non-Commercial License, where commercial use requires a separate written licence from BreezeBlue ^[6]. Shipping commercially requires that licence at any revenue level ^[5].

The dev.to comparison draws the line at monetisation. A free app that generates audio for users and pays for itself with ads is commercial use ^[11]. A university evaluation or benchmark reproduction is fine ^[13]. For an internal tool that reads support tickets aloud to staff inside a revenue-generating business, the post's advice is to talk to BreezeBlue rather than assume ^[12]. The post's own summary is that the licence is what counts: a model that wins on the scoreboard, then disqualifies most of the people reading it ^[17].

Both figures in the price comparison are hosted prices on the same board ^[2]. The gap between them is $66 per 1M characters, which puts the Breeze listing at 34 percent of Eleven v3's rate ^[1]. The route with no per-character bill is the other one: your own GPU, with at least 12 GB of memory, running weights you may not bill anyone for ^[14]. The post does not name the provider serving Breeze at $34, so a buyer reading that row still has to establish whose licence covers the output.

Elo on the Provider Voice Arena is an aggregate of listener votes. For the number to transfer, your text has to resemble what the voters heard and your users have to share their preferences. The post supplies its own error bar: "Elo drifts as votes accumulate, so treat single-digit gaps as ties" ^[10]. Apply that and the five points between Breeze and Eleven v3 Conversational at 1210 is a tie ^[7][2]. Breeze is 40 points clear of Eleven v3 ^[4] and 67 points behind Cartesia Sonic 3.6 ^[3]. It sits seventh of the 98 listed models and first of the 16 open-weight ones ^[3][9]; the next open-weight entry is Fish Audio S2 Pro at 1128 ^[19].

For the low-latency agent case the post still recommends ElevenLabs, and Eleven v3 Conversational lists at $50 per 1M characters, half the flagship rate ^[7][6].

The craft in the release is in the controls. Per the model card and repository README, Breeze TTS 2 offers voice design from a plain-English prompt, voice direction that steers tone, emotion and pace separately from a cloned identity, and inline markers for vocal events such as laughter ^[16]. Designing a voice from a prompt removes the reference clip, and with it the question of whose voice was sampled ^[16]. That is worth the download for evaluation work, which is the work the licence permits.

For a team shipping to paying customers, the options are negotiating terms with BreezeBlue, paying ElevenLabs $50 or $100 per 1M characters, or self-hosting weights that are permissively licensed. VoxCPM2's weights ship under Apache 2.0 ^[15]. I would read the weights licence before the rank table, in that order.