Breeze TTS 2: The New Open Weights TTS Leader at 1,215 Elo

Breeze TTS 2 is now the top-ranked open-weights text-to-speech model on Artificial Analysis's Provider Voices Speech Arena, holding 1,215 Elo — 90 points clear of the previous open-weights leader, Fish Audio S2 Pro, and ranked #6 overall out of more than 100 rated systems. The listing is the first independent confirmation of what BreezeBlue has claimed since it unveiled the model in mid-August 2026: that a two-week-old voice model from a roughly 15-person startup can sit with the commercial text-to-speech frontier. The open weights landed on Hugging Face on August 25, 2026; the arena result followed within a day. Whatever else Breeze TTS 2 turns out to be, it is no longer a vendor promise — it has an Elo that says so.

What actually happened, in order

The timeline matters here, because it explains why an "open weights leaderboard" post is coming from a company most people had not heard of three weeks ago. BreezeBlue was founded by Yang Bin, a former technical co-founder of one of China's leading AI labs, on a $6 million seed round led by Yuanjing Capital and Redpoint China. The model moved through three visible milestones:

• August 7, 2026 — BreezeBlue published its own text-to-speech benchmark suite for voice design, voice direction, and latency evaluation on Hugging Face.

• August 16–17, 2026 — the company formally announced Breeze TTS 2 as a real-time flagship voice model for interactive media, and opened a hosted API at $34 per 1M characters.

• August 25, 2026 — the weights and PyTorch inference code were open-sourced on Hugging Face as BreezeBlue/Breeze-TTS-2, a 3B-parameter model under a research-and-non-commercial license.

The Artificial Analysis Provider Voices ranking was live within days of the open-weights release. Because the arena rates models on blind side-by-side listening, the 1,215 Elo is not a benchmark BreezeBlue ran on itself — it is the accumulated judgement of listeners comparing real samples, which is exactly what a model this young lacks most.

The scoreboard

• Arena rank — #6 overall on Provider Voices; #1 among all open-weights models, ahead of Fish Audio S2 Pro (1,125 Elo).

• Elo — 1,215, with a 95% CI of roughly ±17 (Artificial Analysis, as of late August 2026).

• Languages — 50, per BreezeBlue; the model card is tagged English and Chinese.

• Price — $34 per 1M characters on BreezeBlue's hosted endpoint; open weights are free to download but carry a non-commercial license.

• Speed — ~45 characters per second on Artificial Analysis's measurements — slower than several rivals, which matters for long-form jobs.

• Latency — sub-40ms streaming latency, per BreezeBlue (vendor-stated, not independently measured).

What Breeze TTS 2 actually does differently

The Elo gets the headline, but the product claim is more interesting than the score. Breeze TTS 2 is built around two ideas that most text-to-speech models treat as afterthoughts: voice design and voice direction. Voice design is zero-shot and reference-free — you describe a voice in natural language ("a warm, mid-40s narrator with a slight southern drawl and a dry wit") and the model generates that voice without a studio recording or a reference clip. Voice direction decouples the speaker's timbre from the delivery intent, so you can push a line to sound sarcastic, hushed, or rushed without the model drifting into a different character. BreezeBlue says the same speaker identity survives the shift, and its own voice-direction benchmark reports a speaker similarity of 0.67 SPK_SIM (vendor-reported).

Those two capabilities point at Breeze TTS 2's intended market. The press materials are explicit: video game characters, digital companions, narrative storytelling, broadcast drama, and virtual streamers — long-form, role-driven, multi-character audio, not just another narration API. The company ships a companion voice library called Voice Galaxy with more than 15,000 community and studio-crafted character voices, and BreezeBlue's demo reel is a multi-character fan radio drama that runs over ten minutes. On its own released benchmarks (vendor-reported, unreproduced), Breeze TTS 2 claims the #1 spot on the text-to-speech Voice Design benchmark with a Role Fit score of 78.02 against 72.78 for MiMo-V2.5-TTS, and a Voice Direction composite of 4.25.

The license asterisk

Before the "open weights" phrase does too much marketing work, read the model card. Breeze TTS 2 is 3B parameters, safetensors, F32/BF16, and the source code is Apache 2.0 — but the weights, derivative models, and self-hosted outputs are licensed for research and non-commercial use only, under the breezeblue-research-and-non-commercial-license. That means "open weights" here is not "free for your product." Commercial self-hosting is not permitted by the license as written; the commercial path is the hosted API at $34 per 1M characters. This is the same shape as several other 2026 open releases — inspectable and self-hostable for research, but with the revenue-bearing use reserved for the vendor's own endpoint.

Why a router matters for a model this young

The honest risk with Breeze TTS 2 right now is not quality — it is age. The model has been live for ten days, the weights for two. No production team should bet a voice pipeline on a model that young without a way to move off it, and that is precisely the failure mode a routing layer exists to absorb. If you evaluate Breeze TTS 2 through a gateway that treats the vendor endpoint as one route among many, you get the Elo leader today, automatic failover to a fallback provider when the API is degraded, and a one-line change if a week-old model shows a regression. That is the same discipline the routing DSL applies to any model: compose it with alternatives, keep the interface stable, and let the model prove itself before you let it become load-bearing. None of the models in this series is yet routed on OrcaRouter itself, so the honest advice is to wire Breeze TTS 2 into whatever gateway you already run — and keep your current provider live alongside it while the Elo gets older and more trustworthy.

What to watch next

The Provider Voices #1-open-weights slot is the story today, but it is the weaker of the two arenas for this model. On Artificial Analysis's Controlled Voice arena, where all models are compared on the same fixed reference voices, Breeze TTS 2 ranks #3 among open-weights models at 1,002 Elo, behind Mistral's Voxtral at 1,010 — and #16 overall out of 39. That gap between its provider-voice score (1,215, #1 open) and its controlled-voice score (1,002, #3 open) is worth watching: it suggests Breeze TTS 2's advantage is concentrated in the voices it designs for itself, which is exactly the product claim, and it is also exactly the claim an independent benchmark should keep pressure on. Watch whether the controlled-voice Elo closes toward the provider-voice number, whether the commercial license tightens or loosens, and — most of all — whether anyone reproduces the voice-design and voice-direction benchmarks outside BreezeBlue's own suite.

For a model that did not exist a month ago, Breeze TTS 2 has already earned the most useful thing a text-to-speech model can earn: an independent score that agrees with the marketing. Whether it becomes a default voice for interactive products depends less on the next Elo update than on how the license and the self-hosting story evolve — but as of this week, open-weights text-to-speech has a new leader, and it is two weeks old.